Hyphae are polarized, elongating and bifurcating cellular structures that many fungi use to forage and feed (figure 1a and b). The phylum Chytridiomycota (chytrids) diverged from other fungal lineages approximately 750 Mya and, with the Blastocladiomycota, formed a critical evolutionary transition in the Kingdom Fungi dedicated to osmotrophy and the establishment of the chitin-containing cell wall [2]. Chytrids produce filamentous hyphae-like, anucleate structures called rhizoids (figure 1a–c) [3], which are important in their ecological functions, in terms of both attachment to substrates and osmotrophic feeding [2]. 407-million-year-old fossils from the Devonian Rhynie Chert deposit show chytrids in freshwater aquatic ecosystems physically interacting with substrates via rhizoids in a comparative mode to extant taxa [4]. Yet surprisingly, given the importance of rhizoids in both contemporary and paleo-chytrid ecology, there remains a limited understanding of chytrid rhizoid biology, including possible similarities with functionally analogous hyphae in other fungi and the potential for substrate-dependent adaptations. Show  Figure 1. Rhizoids are the basal feeding condition within the fungal kingdom and their morphogenesis is similar to hyphal development. (a–b) Correlating the major feeding types in fungi (a) to phylogeny (b) shows rhizoids to be the basal feeding condition in the true fungi (Eumycota). White circles indicate absence of a growth plan in a taxon and dark circles indicate widespread presence. Faded circles indicate a growth plan is present within a taxon, but not widespread. Rhizoid systems bear a single thallus, whereas rhizomycelia are defined as rhizoid systems with multiple thalli. Tree based on the phylogeny outlined in [1]. (c) R. globosum displays a typical chytrid cell plan. Shown is the thallus (th) anchored to the substrate by threadlike rhizoids. Rhizoids emanate from a swelling termed the apophysis (a). Also shown are rhizoid bifurcations (rb) and tips (rt). Scale bar = 5 µm. (d) R. globosum exhibits an archetypal chytrid life cycle. From the mature zoosporangium, multiple motile zoospores are released that swim freely for a brief period (less than 1–2 h) before encysting into germlings (i.e. losing the single flagellum and developing a chitin-containing cell wall). From the subsequent extending germtube develops the rhizoid. After a period of growth, the sporangium becomes the multinucleate zoosporangium during sporogenesis, from which the next generation of spores is released. (e) Chytrid rhizoids were reconstructed using the neuron tracing workflow outlined in electronic supplementary material, figure S3. Example of a three-dimensional reconstructed R. globosum rhizoid system taken from a 10 h time series. Scale bar = 20 µm. (f) This study used morphometric parameters developed from neuron biology to described rhizoid development in chytrids. Shown are the Euclidean distance (Eucl. Dist.), bifurcation angle (Bif. Angle), rhizoid growth unit (RGU) and cover area. Full morphometric glossary is presented in electronic supplementary material, figure S5. (g) Rhizoid growth trajectories for four-dimensional confocal time series (n = 5, mean ± s.e.m.) of rhizoidal growth unit, total length and number of tips. (h) Apical and lateral branches occur in chytrid rhizoids. Apical branching occurs when a branch is formed at the rhizoid tip parallel to the established rhizoidal axis. Lateral branching occurs when a branch is formed distally to the rhizoidal tip, establishing a new rhizoidal axis. (i) Four-dimensional confocal imaging (n = 5, mean ± s.e.m.) revealed that lateral branching dominates over apical branching *p < 0.05. (j) Fractal analysis of chytrid rhizoid systems shows a decrease in fractal dimension (Db) towards the growing edge. Thallus demarked by dashed circle. Scale bar = 50 µm.
Character mapping of the presence of cellular growth plans against established phylogenies reveals the multicellular hyphal form to be a derived condition, whereas rhizoid feeding structures are the basal condition within the true fungi (Eumycota) (figure 1a and b). Aseptate hyphae represent an intermediary condition and are not typically the dominant cell type in either unicellular or multicellular fungi. Hyphal cell types are sometimes observed outside of the Eumycota, such as within the Oomycota; however, the origin of fungal hyphae within the Eumycota was independent [5,6] and has not been reported in their closest relatives the Holozoans (animals, choanoflagellates, etc.). Comparative genomics has indicated that hyphae originated within the rhizoid-bearing Chytridiomycota–Blastocladiomycota–Zoopagomycota nodes of the fungal tree [6], which is supported by fossil Blastocladiomycota and extant Monoblepharidomycetes having hyphae [5,7]. This has led to the proposition that rhizoids, or rhizoid-like structures, were the evolutionary precursors of fungal hyphae [5,6,8]; however, investigation into such hypotheses have been hindered by a relative lack of understanding of rhizoid developmental biology. Chytrids are important aquatic fungi [9], feeding on a range of physically complex heterologous substrates, including algal cells [10], amphibian epidermises [11] and recalcitrant particulate organic matter (POM) such as chitin and pollen [12]. Appreciation for the ecological importance of chytrids as parasites, pathogens and saprotrophs in aquatic ecosystems is greatly expanding. For example, chytrids are well-established plankton parasites [10], responsible for the global-scale amphibian panzootic [13], and have recently emerged as important components of the marine mycobiome [9]. The chytrid rhizoid is critical in all ecological functions because it is the physical interface between the fungus and substrate or host, yet there remains a limited understanding of rhizoid functional biology in terms of substrate interaction. Rhizoclosmatium globosum is a widespread aquatic saprotroph that is characteristically associated with chitin-rich insect exuviae and has an archetypal chytrid cell plan (figure 1c) and life cycle (figure 1d) [14]. With an available sequenced genome [15], easy laboratory culture and amenability to live cell imaging (this study), R. globosum JEL800 represents a promising model organism to investigate the cell and developmental biology of aquatic rhizoid-bearing, early-diverging fungi. To study the developing rhizoid system for morphometric analyses, we established a live cell three-/four-dimensional confocal microscopy approach in combination with the application of neuron tracing software to three-dimensional reconstruct developing cells (figure 1e; electronic supplementary material, figures S3 and S4). We were subsequently able to generate a series of cell morphometrics to describe and quantify rhizoid development (figure 1f; electronic supplementary material, figure S5) under a range of experimental conditions with the aims of identifying potential similarities with hyphae in dikaryan fungi in terms of geometric organization, morphogenesis and underlying cellular control mechanisms. In addition, we set out to characterize substrate-dependent adaptations particularly in the ecological context of aquatic POM utilization. Detailed materials and methods are provided as electronic supplementary material. Chytrid plasma membranes were labelled with 8.18 µM FM 1–43 and imaged using a Zeiss LSM 510 Meta confocal laser scanning microscope (Carl Zeiss). Z-stacks of rhizoids were imported into the neuron reconstruction software NeuronStudio [16,17] and semi-automatically traced with the ‘Build Neurite’ function. Rhizoids were morphometrically quantified using the btmorph2 library [18] run with Python 3.6.5 implemented in Jupyter Notebook 4.4.0. For visualization, reconstructed rhizoids were imported into Blender (2.79), smoothed using automatic default parameters and rendered for display only. To label F-actin and the cell wall throughout the rhizoid system, cells were fixed for 1 h in 4% formaldehyde in 1×PBS (phosphate-buffered saline) and stained with 1 : 50 rhodamine phalloidin in PEM (100 mM PIPES (piperazine-N,N′-bis(2-ethanesulfonic acid)) buffer at pH 6.9, 1 mM EGTA (ethylene glycol tetraacetic acid), and 0.1 mM MgSO4) for 30 min, then with 5 µg ml−1 Texas Red-conjugated wheat germ agglutinin (WGA) in PEM for 30 min. Caspofungin diacetate (working concentration 1–50 µM) was used to inhibit cell wall β-glucan synthesis and cytochalasin B (working concentration 0.1–10 µM) was used to inhibit actin filament formation. Cells were incubated for 6 h, which was found to be sufficient to observe phenotypic variation. R. globosum was processed for β-glucans using a commercial β-glucan assay (Yeast & Mushroom) (K-YBGL, Megazyme) following the manufacturer's protocol. Briefly, samples were processed by acid hydrolysis then enzymatic break-down and β-glucans were quantified spectrophotometrically with a CLARIOstar Plus microplate reader (BMG Labtech), relative to a negative control and positive β-glucan standard. A sample of shop-bought baker's yeast was used as an additional positive control. All glycosyl transferase group 2 (GT2) domain-containing proteins within the R. globosum genome were identified using the JGI MycoCosm online portal. GT2 functional domains were identified using DELTA-BLAST [19] and aligned with MAFFT [20]. Maximum-likelihood phylogenies were calculated with RAxML [19] using the BLOSUM62 matrix and 100 bootstrap replicates. For carbon starvation experiments, R. globosum cells were grown in either carbon-free Bold's Basal Medium supplemented with 1.89 mM ammonium sulfate and 500 µl l−1 f/2 vitamin solution [21] (BBM) or BBM with 10 mM N-acetyl-D-glucosamine as a carbon source. To investigate growth on POM, chitin microbeads (New England Biolabs) were suspended in carbon-free BBM at a working concentration of 1 : 1000 stock concentration. To understand rhizoid development in a starved cell that had encountered a chitin microbead, we imaged cells that contacted a chitin microbead following development along the glass bottom of the dish. During rhizoid development, we observed a continuous increase in rhizoid length (110.8 ± 24.4 µm h−1) (n = 5, ± s.d.) and the number of rhizoid tips (4.6 ± 1.2 tips h−1) (figure 1g; electronic supplementary material, table S1, movies S1–S5), with an increase in the total cell surface area (21.1 ± 5.2 µm2 h−1), rhizoid bifurcations (4.2 ± 1.0 bifurcations h−1), cover area (2,235 ± 170.8 µm2 h−1) and maximum Euclidean distance (5.4 ± 0.1 µm h−1) (electronic supplementary material, figure S6). The hyphal growth unit (HGU) has been used previously to describe hyphal development in dikaryan fungi and is defined as the distance between two hyphal bifurcations [22]. Adapting this metric for the chytrid rhizoid, the rhizoidal growth unit (RGU) (i.e. the distance between two rhizoid bifurcations; figure 1f) increased continuously during the first 6 h of the development period (i.e. cells became relatively less branched) before stabilizing during the later phase of growth (figure 1g). The local rhizoid bifurcation angle remained consistent at 81.4° ± 6.3 after approximately 2 h (electronic supplementary material, figure S6), and lateral branching was more frequent than apical branching during rhizoid development (figure 1h and i). Fractal analysis (fractal dimension = Db) of 24 h grown cells showed that rhizoids approximate a two-dimensional biological fractal (mean Db = 1.51 ± 0.24), with rhizoids relatively more fractal at the centre of the cell (max Db = 1.69–2.19) and less fractal towards the growing periphery (min Db = 0.69–1.49) (figure 1j; electronic supplementary material, figure S7). The cell wall and actin structures were present throughout the chytrid rhizoid (figure 2a). Putative actin cables ran through the rhizoid system, punctuated by actin patches. Inhibition of cell wall β-1,3-glucan synthesis and actin proliferation with caspofungin and cytochalasin B, respectively, induced a concentration-dependent decrease in the RGU, with the development of atypical hyperbranched rhizoids (figure 2b–d; electronic supplementary material, table S2, movies S6–S7). As with Batrachochytrium dendrobatidis [23,24], we confirmed that R. globosum JEL800 lacks an apparent β-1,3-glucan synthase FKS1 gene homologue (electronic supplementary material, table S3). However, the quantification of glucans in R. globosum showed that they are present (figure 2e), with 58.3 ± 7.6% β-glucans and 41.6 ± 7.6% α-glucans of total glucans. Figure 2. Cell wall synthesis and actin dynamics govern rhizoid branching. (a) Fluorescent labelling of cell wall and actin structures in 24 h R. globosum rhizoids. The cell wall and actin patches were found throughout the rhizoid. Arrowheads in the actin channel indicate putative actin cables. WGA = conjugated wheat germ agglutinin. Scale bar = 10 µm. (b) Representative three-dimensional reconstructions of 7 h R. globosum cells following treatment with caspofungin diacetate and cytochalasin B at stated concentrations to inhibit cell wall and actin filament biosynthesis respectively, relative to solvent only controls. Scale bar = 20 µm. (c) Application of caspofungin diacetate and cytochalasin B resulted in a concentration-dependent decrease in the rhizoidal growth unit, resulting in atypical hyperbranched rhizoids (n ∼ 9). n.s p > 0.05 (not significant), *p < 0.05, **p < 0.01, ***p < 0.001. This differential growth is diagrammatically summarized in (d). (e) β-glucan concentration of R. globosum (n = 10) relative to a baker's yeast control (n = 2). (f) Maximum-likelihood phylogeny of GT2 domains (BcsA and WcaA domains) within the R. globosum genome (midpoint rooting). Full architecture of each protein is shown. Asterisk indicates the putative glucan synthesis protein ORY39038 containing a putative SKN1 domain.
To identify alternative putative β-glucan synthesis genes in R. globosum JEL800, we surveyed the genome and focused on GT2 encoding genes, which include typical glucan synthases in fungi. A total of 28 GT2 domains were found within 27 genes (figure 2f). Of these genes, 20 contained putative chitin synthase domains and many contained additional domains involved in transcriptional regulation. Nine encode chitin synthase 2 family proteins and 11 encode chitin synthase 1 family proteins (with two GT2 domains in ORY48846). No obvious genes for β-1,3-glucan or β-1,6-glucan synthases were found within the genome. However, the chitin synthase 2 gene ORY39038 included a putative SKN1 domain (figure 2f), which has been implicated in β-1,6-glucan synthesis in the ascomycete yeasts Saccharomyces cerevisiae [25] and Candida albicans [26]. These results indicate a yet uncharacterized β-glucan-dependent cell wall production process in chytrids (also targeted by caspofungin) that is not currently apparent using gene/genome level assessment and warrants further study. To examine whether chytrids are capable of modifying rhizoid development in response to changes in resource availability, we exposed R. globosum to carbon starvation (i.e. development in the absence of exogenous carbon). When provided with 10 mM N-acetylglucosamine (NAG) as an exogenous carbon source, the entire life cycle from zoospore to sporulation was completed and the rhizoids branched densely (electronic supplementary material, movie S8), indicative of a feeding phenotype. Carbon-starved cells did not produce zoospores and cell growth stopped after 14–16 h (electronic supplementary material, movie S9). Using only endogenous carbon (i.e. zoospore storage lipids), starved cells underwent differential rhizoid development compared to cells from the exogenous carbon-replete conditions to form an apparent adaptive searching phenotype (figure 3a,b; electronic supplementary material, table S4, movie S10). Under carbon starvation, R. globosum invested less in thallus growth than in carbon replete conditions and developed longer rhizoids with a greater maximum Euclidean distance (figure 3c). Carbon-starved cells were also less branched, had wider bifurcation angles and subsequently covered a larger surface area. These morphological changes in response to exogenous carbon starvation (summarized in figure 3b) suggest that individual chytrid cells are capable of differential reallocation of resources away from reproduction (i.e. the production of the zoosporangium) and towards an extended modified rhizoidal structure indicative of a resource searching phenotype. Figure 3. Chytrids are capable of adaptive rhizoid development under carbon starvation. (a) Representative three-dimensional reconstructions of R. globosum cells grown under carbon-replete or carbon-depleted conditions at different time points. Scale bar = 20 µm. When exposed to carbon starvation, chytrids are capable of differential adaptive growth to produce a searching phenotype. This differential growth is summarized in (b). (c) Differential growth trajectories of major morphometric traits between R. globosum cells (n ∼9, mean ± s.e.m.) grown under carbon-replete and carbon-depleted conditions over time. n.s p > 0.05 (not significant), *p < 0.05, **p < 0.01, ***p < 0.001.
Rhizoid growth of single cells growing on chitin microbeads was quantified as experimental POM (figure 4a,b; electronic supplementary material, movie S11). Initially, rhizoids grew along the outer surface of the bead and were probably used primarily for anchorage to the substrate. Scanning electron microscopy (SEM) showed that the rhizoids growing externally on the chitin particle formed grooves on the bead parallel to the rhizoid axis (electronic supplementary material, figure S1f,g), suggesting extracellular enzymatic chitin degradation by the rhizoid on the outer surface. Penetration of the bead occurred during the later stages of particle colonization (figure 4a; electronic supplementary material, movie S12). Branching inside the bead emanated from ‘pioneer’ rhizoids that penetrated the particle (figure 4c). Figure 4. Rhizoids associated with heterogenous particulate carbon exhibit spatial differentiation. (a) Representative three-dimensional reconstructions of R. globosum cells (blue) growing on chitin microbeads (beige) at different time points. Scale bar = 20 µm. (b) Growth trajectories for total rhizoid length and thallus surface area for R. globosum cells growing on chitin microbeads (n ∼9, mean ± s.e.m.). (c) Diagrammatic summary of R. globosum rhizoid development on chitin microbeads. (d) Representative three-dimensional reconstruction of a 24 h searching R. globosum cell (blue) that has encountered a chitin microbead (beige). The colour coded panel shows parts of the rhizoid system in contact (green) and not in contact (blue) with the microbead. Scale bar = 20 µm. (e) Comparison of rhizoids in contact or not in contact with the chitin microbead (n = 8). (f) Diagrammatic summary of spatial differentiation in a starved, searching rhizoid that has encountered a particulate carbon patch.
Given the previous results of the searching rhizoid development in response to carbon starvation, a patchy resource environment was created using the chitin microbeads randomly distributed around individual developing cells in otherwise carbon-free media to investigate how encountering POM affected rhizoid morphology (figure 4d). We observed spatial differentiation of single-cell rhizoid systems in association with POM contact. Particle-associated rhizoids were shorter than rhizoids not in particle contact, were more branched (i.e. lower RGU), had a shorter maximum Euclidean distance and covered a smaller area (figure 4e). These rhizoid morphometrics closely resembled the feeding and searching modifications of the cells grown under carbon-replete and carbon-depleted conditions previously discussed (figures 4f and 3b) but instead are displayed simultaneously with spatial regulation in individual cells linked to POM-associated and non-associated rhizoids, respectively. Our results provide new insights into the developmental cell biology of chytrid fungi and highlight similarities between the organization of anucleate rhizoids and multicellular hyphae. The fundamental patterns of rhizoid morphogenesis that we report here for a unicellular non-hyphal fungus are comparable to those previously recorded for hyphal fungi (figure 5a) [22]. Trinci [22] assessed hyphal development in major fungal lineages (Ascomycota and Mucoromycota) and observed that the growth patterns of morphometric traits (HGU, total length and number of tips) were similar across the studied taxa. When the data from our study are directly compared to that of Trinci [22], we see that the hyphal growth pattern is also analogous to the rhizoids of the early-diverging unicellular Chytridiomycota (figure 5a). Chytrid rhizoid development in this study is also comparable to the hyphal growth rates reported by Trinci [22] (figure 5b,c), as well as the elongation rates reported by López-Franco et al. [27] when scaled by filament diameter (figure 5d). Figure 5. Development of chytrid rhizoids fundamentally resembles mycelial development in hyphal fungi. Comparison of rhizoid development from this study (asterisks) with other studies on hyphal fungi [22,27]. (a) Growth trajectories of the growth unit, total length and number of tips of rhizoids and hyphae. Data for other fungi are reproduced as new figures directly from [22]. For R. globosumn = 5 biological replicates. Error bars denote ± s.e.m. (b,c) R. globosum has similar growth rates regarding total length (b) and tip production (c) to hyphal fungi. Rhizoid growth rates (µ) calculated as increase in the total rhizoid length or tip number as in [22]. Data for Ascomycota and Mucoromycota fungi are from [22]. For R. globosum n = 5 biological replicates. Mean ± s.d. (d) R. globosum rhizoids when scaled by diameter have a comparable elognation rate to hyphae. Rhizoid elongation rates (the speed at which individual rhizoid compartments extend) were quantified by measurement of extending rhizoid compartments (10 rhizoids for each biological replicate) separated by a 30 min interval on maximum intensity projected z-stacks in Fiji. Data for other fungi are from [27]. For R. globosum n = 5 biological replicates. Mean ± s.e.m.
Such similarities also extend to rhizoid branching patterns, where lateral branching dominates over apical branching. This branching pattern is also the predominant mode of hyphal branching, where apical branching is suppressed by a phenomenon termed ‘apical dominance’ [28]. These findings suggest that a form of apical dominance at the growing edge rhizoid tips may suppress apical branching to maintain rhizoid network integrity as in dikaryan hyphae [28,29]. Chytrid rhizoids also become less fractal towards the growing edge in terms of their overall morphology, and similar patterns of fractal organization are also observed in hyphae-based mycelial colonies [30]. Taken together, these results show strong geometric analogies in the fundamental organization of unicellular chytrid rhizoid and multicellular hyphal morphogenesis. Given the apparent hyphal-like properties of rhizoid development, we sought a greater understanding of the potential subcellular machinery underpinning rhizoid morphogenesis in R. globosum. Normal rhizoid branching was disrupted by inhibition of cell wall synthesis and actin polymerization, both of which are known to control branching and growth in hyphal biology [31–33]. These effects in R. globosum are similar to disruption of normal hyphal branching reported in Aspergillus fumigatus (Ascomycota) in the presence of caspofungin [34], and in Neurospora crassa (Ascomycota) in the presence of cytochalasins [35]. Recent studies have shown the presence of actin in the rhizoids of soil chytrids [36,37] and inhibition of actin in Chytriomyces hyalinus similarly disrupts normal rhizoid branching [36]. In this study, our quantitative characterization of cell wall and actin inhibited rhizoid paramorphs provides support that β-1,3-glucan-dependent cell wall synthesis and actin dynamics also govern branching in chytrid rhizoids as in multicellular hyphae. We also show that rhizoid development is plastic to resource availability, with chytrid cells displaying an adaptive searching phenotype under carbon starvation. Adaptive foraging strategies are well described in multicellular hyphae [38,39], and our data support the existence of analogous strategies in rhizoidal fungi. Dense branching zones in dikaryan mycelia are known to improve colonization of trophic substrates and feeding by increasing surface area for osmotrophy, while more linear ‘exploring’ zones cover greater area and search for new resources [39]. Similar morphometrics are displayed by R. globosum exhibiting feeding and searching phenotypes, respectively. In addition, exogenous carbon starvation has also been shown to be associated with a decrease in branching in the multicellular dikaryan fungus Aspergillus oryzae (Ascomycota) [40]. Overall, these results highlight that adaptive search strategies are more widely spread than previously known in the Kingdom Fungi. Finally, we report the spatial and functional differentiation of feeding and searching sections of anucleate rhizoid systems from individual cells. The simultaneous display of both rhizoid types in the same cell indicates a controlled spatial regulation of branching and differentiation of labour within single chytrid rhizoid networks. Functional division of labour is prevalently seen in multicellular mycelial fungi [38,39] including the development of specialized branching structures for increased surface area and nutrient uptake, as in the plant symbiont mycorrhiza (Glomeromycota) [41]. Our observation of similarly complex development in a unicellular chytrid suggests that multicellularity is not a prerequisite for adaptive spatial differentiation in fungi. The improved understanding of chytrid rhizoid biology related to substrate attachment and feeding we present here opens the door to a greater insight into the functional ecology of chytrids and their environmental potency. Our approach of combining live cell confocal microscopy with three-dimensional rhizoid reconstruction provides a powerful toolkit for morphometric quantification of chytrid cell development and could shed light on the biology underpinning chytrid ecological prevalence. In the future, the application of this approach to different systems could provide a detailed understanding of chytrid parasitism and host interaction, development under different nutrient regimes and degradation of diverse carbon sources. From an evolutionary perspective, the early-diverging fungi are a critical component of the eukaryotic tree of life [42,43], including an origin of multicellularity and the establishment of the archetypal fungal hyphal form, which is responsible in part for the subsequent colonization of land by fungi, diversity expansion and interaction with plants [2]. Our cell biology focused approach advances this developing paradigm by showing that a representative unicellular, rhizoid-bearing (i.e. non-hyphal) chytrid displays hyphal-like morphogenesis, with evidence that the cell structuring mechanisms (e.g. apical dominance) underpinning chytrid rhizoid development are equivalent to reciprocal mechanisms in dikaryan fungi. Perhaps our key discovery is that the anucleate chytrid rhizoid shows considerable developmental plasticity. R. globosum is able to control rhizoid morphogenesis to produce a searching form in response to carbon starvation and, from an individual cell, is capable of spatial differentiation in adaptation to patchy substrate availability indicating functional division of labour. The potential for convergent evolution aside, we propose by parsimony from the presence of analogous complex cell developmental features in an extant representative chytrid and dikaryan fungi that adaptive rhizoids are a shared feature of their most recent common ancestor. All data that support the findings of this study are included in the electronic supplementary material of this paper. D.L. and M.C. conceived the study. D.L. conducted the laboratory work and data analysis. N.C. analysed the R. globosum JEL800 genome. G.W. provided support with microscopy. M.C. secured the funding. D.L. and M.C. critically assessed and interpreted the findings. D.L and M.C. wrote the manuscript, with the help of N.C. and G.W. The authors declare no competing interests. D.L. is supported by an EnvEast Doctoral Training Partnership (DTP) PhD studentship funded from the UK Natural Environment Research Council (NERC grant no. NE/L002582/1). M.C. is supported by the European Research Council (ERC) (MYCO-CARB project; ERC grant agreement no. 772584). N.C. is supported by NERC (Marine-DNA project; NERC grant no. NE/N006151/1). G.W. is supported by an MBA Senior Research Fellowship. The authors would like to thank Glenn Harper, Alex Strachan and the team at the Plymouth Electron Microscopy Centre (PEMC) for their assistance. We are indebted to Joyce Longcore (University of Maine) for providing R. globosum JEL800 from her chytrid culture collection (now curated by the Collection of Zoosporic Eufungi at the University of Michigan). FootnotesElectronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.5001059. © 2020 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original author and source are credited. References
Page 2Plague (caused by infection with Yersinia pestis) is a historical bacterial disease which remains a substantial concern to global public health [1]. In the past decade, plague outbreaks have been reported in Madagascar [2], the Democratic Republic of Congo [3] and Peru [4], and the bacterium is regularly detected among different small rodent reservoirs in the USA [5], China [6] and Kazakhstan [7]. These small rodents have the potential to spread the disease to humans via bites from infected fleas [8]. Infected humans will develop tender swollen lymph nodes, known as buboes [9], from which the infection may later spread to the lungs via the bloodstream leading to secondary pneumonic plague [10]. At this point, the bacteria can be directly transmitted from human to human through respiratory droplets [11]. Several pathways within this transmission cycle are influenced by climate [12]. Cold and dry environments can hinder the survival and development rates of flea eggs and larvae [13,14]. Ambient temperatures that are too low or too high can also inhibit flea-gut blockage [15–17]—a proposed mechanism of successful bacterial transmission [18]. However, if blockage has already occurred, the role of temperature on bacterial survival could be more crucial in determining transmission efficiency [19]. The population dynamics and behaviour of rodent hosts are also affected by seasonal variation in temperature and precipitation [20]. High rainfall can flood rodent burrows, driving them towards urban areas [21] and low resources during winter can reduce rodent populations [22]. Combining these factors can have drastic effects on plague epidemiology [23–25]; climate drivers can facilitate the introduction of bacteria into naive rodent populations [26,27], from which plague outbreaks can emerge annually within human populations [9,28]. In the absence of an effective licensed vaccine [29], it is important to understand the epidemiological drivers of plague to inform public health planning. Detailed historical records offer a valuable opportunity to study long-term plague on a large spatio-temporal scale given the scarcity of modern-day outbreaks. The extent of the third plague pandemic, originating in Yuhan Province, China around 1855 [30] before spreading globally throughout the nineteenth and twentieth centuries [31], allows us to rigorously quantify the effects of climate on plague epidemiology across different spatial contexts. The effects of climate on the annual cases of plague in China [32] and pre-industrial Europe [33,34] have already been demonstrated; here, we consider the dynamics at a finer scale, and present a 50 year historical dataset of monthly provincial plague-related deaths during the third plague pandemic in British India, one of the most severely affected regions during the third plague pandemic [35,36], including modern-day Pakistan, India, Bangladesh and Myanmar. We analyse how plague emerged annually throughout the region from 1898 to 1949 and, using temperature, rainfall and humidity data from the same time-period, we show the role of climate on the likelihood of outbreaks occurring. Finally, we demonstrate the relationship between the timing of annual plague outbreaks throughout British India against seasonal climate variation. After the introduction of plague into British India in 1896, data for monthly plague deaths per province in India were available throughout 1898–1949 in the annual reports of the Chief Sanitary Commissioner of India [36]. The reports also contained monthly plague deaths for Bombay City (Mumbai), Madras City (Chennai), Calcutta City (Kolkata) and Bangalore Civil and Military Station. For Upper and Lower Burma, data were only available from 1904 to 1922; other years contained spatially aggregated averages for the whole of Burma. Provinces constituted parts of modern-day Pakistan, India, Bangladesh and Myanmar. Each location was indexed from the north, starting with North-West Frontier Province, to the southern tip of the Indian Peninsula. The location of each province and corresponding identification number are shown in figure 1. Figure 1. Provinces in British India. The map shows the location of each province identified in the plague data from the annual Chief Commissioner reports. The data encapsulated modern-day Pakistan, India, Bangladesh and Myanmar. Provinces without any available plague data from 1898 to 1949 are shown in grey.
Population data for each province in India were obtained from 1901, 1911, 1921, 1931 and 1941 census of India [37]. Population sizes between these years were estimated using cubic spline interpolation through this data [38]. The number of plague deaths per 1 000 000 individuals per province from 1898 to 1949 was then calculated using these population estimates. Climate data were obtained from the annual and monthly weather reviews of India provided by the National Oceanic and Atmospheric Administration [39]. Climate records on the same spatio-temporal scale as the plague data were only available for the years between 1907 and 1936. These data contained monthly averages for temperature (in °F), relative humidity and total rainfall (in inches). Monthly weather records were not consistently available for Delhi, Ajmer Merwara, Coorg, Bombay City, Bangalore Civil and Military Station, Madras City and Calcutta. Humidity data were also not available for Upper and Lower Burma, Jammu and Kashmir and Baluchistan Agency. In the place of Lower and Upper Burma, spatially averaged relative humidity was available for the entirety of Burma. Some records had missing values for monthly climate averages from select provinces. These missing values were replaced by manually averaging across all available data from weather station records found in the appendix of the weather reviews. These climate data were digitized by image processing tables of values in GIMP [40] and performing image recognition in R using the tesseract package [41,42]. In total, 16 837 climate record entries were digitized. All digitized data were then manually checked against the annual and monthly weather reviews and adjusted to match the raw data. Historical maps depicting administrative boundaries in the years 1901, 1911, 1921, 1931 and 1941 were obtained from the Administrative Atlas of India [43]. These maps were then digitized in QGIS [44] down to the provincial level. Missing maps (owing to administrative boundary changes in intermediate years) were then constructed. For each year, plague data were then paired with the map which contained the same provinces. Wavelet analysis [45] was performed on the time-series of monthly plague-related deaths per 1 000 000 individuals for each province using the R package WaveletComp [46]. This approach deconstructed each time-series into a series of Mortlet wavelets of different periodicities [45]. The periodicity of plague outbreaks in each province was then calculated as the mean period of the wavelet with the largest magnitude across all time points. The timing of annual plague outbreaks and oscillations in temperature, rainfall and humidity was then calculated by comparing the phase of annual wavelet components to give the approximate time during the year at which plague outbreaks and fluctuations in climate peaked for each year. To understand the role of climate on the timing of plague outbreaks, time lags between plague outbreaks and climate for each location, denoted here as ψtp/tc(t,l), were calculated as follows: ψtp/tc(t,l)=ϕtp(t,l)−ϕtc(t,l), where ϕp (t, l) and ϕc (t, l) are the phases of the annual wavelet components corresponding to time t and location l for the plague outbreaks and climate time-series, tp and tc, respectively. The mean annual time lag for each province was then calculated by averaging over the time lag across all months in each year.In order to assess the relationship between climate and plague outbreaks in British India during 1898–1949, regression models were fitted to the empirical data within a Bayesian framework. For all models, weakly informative independent Cauchy priors were used for the coefficients, βi, with mean zero and scale parameter equal to 10 and 2.5 for the intercepts and slopes respectively. Exponential priors with rate one were used for variance parameters. For models with multiple predictors, all covariates were orthognalized using QR-decomposition, models fitted and coefficients back-transformed to the scale of the data [47]. All models were fitted using the package rstanarm in R [48] and convergence and fit was assessed through visual inspection of the posterior predictive distribution and Gelman–Rubin statistic, R^ [49]. Bayes factors [50], B, were used to assess the strength of evidence in favour of each model against their respective null model—a model containing no linear covariates. A Bayes factor of greater than one indicated favour towards the model, whereas less than one indicated favour towards the null model. A more detailed interpretation of Bayes factors can be found in the electronic supplementary material, table S1. The R-package bayestestR [51,52] was used to compute Bayes factors via bridge-sampling [53] to estimate the marginal-likelihood of each model (see [54] for more details). Bayes factors were also calculated to determine the relative probability that each model parameter was non-zero, denoted Bβi, using the Savage-Dickey density ratio [55]. Each model was fitted to temperature, rainfall and humidity data separately. All regression model results are presented in the electronic supplementary material, tables S2–S4. An outbreak was said to occur if the total number of plague-related deaths per 1 000 000 individuals within a year was (strictly) more than some outbreak threshold. The following logistic regression model was then fitted to estimate the probability of outbreaks occurring based on annual climate averages: χα∼Binom (1,p),logit (p)=β0+β1x+β2x2, where χα was whether an outbreak occurred or not given an outbreak threshold α, and x denotes annual climate averages. For rainfall, an additional cubic term, β3x3 was added to the model. Outbreak thresholds of 0, 1, 10 and 100 plague-related deaths per 1 000 000 individuals were tested.The magnitude of an outbreak was defined as the total number of plague-related deaths reported during each year where an outbreak had occurred given an outbreak threshold of zero. The effects of temperature, rainfall and humidity on outbreak magnitude were then estimated using the following model: log(y)∼N(μy,σy2),μy=β4+β5x+β6x2, where y was the total number of reported plague deaths per year, x denotes annual climate averages and σy2 is the variance of log (y) to be estimated. Similarly to the above regression model, an additional cubic term, β7x3 was added to the model for rainfall.In order to estimate the relationship between the timing of plague outbreaks and oscillations in climate, the following model was fitted to the time at which plague outbreaks peaked during the year within each province, τp: τp∼N(μτp,στp2),μτp=β8+β9τc, where τc denotes the time at which oscillations in climate peak (calculated from Wavelet analysis), and στp2 is the variance of τp to be estimated. The estimated coefficient β9 represented the additional (on average) time-lag of a plague outbreak given a one-month time-lag in oscillations of each climate variable.Plague was first reported to the Chief Sanitary Commissioner of India during the latter part of 1896 in the west of the Indian Peninsula. By 1898, plague-related deaths were reported throughout British India. Annual deaths within each province increased until 1905, when over 22 of the 25 provinces experienced over 100 deaths per one million individuals (figure 2). The size of annual outbreaks in each province then decreased until 1930 to around five deaths per one million individuals. From there on, a low level of background transmission was maintained until 1950. Over 13 million plague-related deaths in total were reported across British India from 1898 to 1949. Figure 2. Annual plague deaths in each province of British India from 1898 to 1949. The total number of plague deaths showed large variation between each province within the same year. When the third pandemic began in around 1898, the total number of plague-related deaths per province rapidly increased each year. From 1905 onwards, plague-related deaths steadily decreased to low levels during the 1940s. (Online version in colour.)
During this period, the north, northwest and west of the Indian Peninsula were the most severely affected. These regions included Bombay Presidency and Punjab provinces, which reported over two and three million cases, respectively. By contrast, cases were infrequently reported in the drier far northwest and wetter northeast of British India, including the provinces Baluchistan Agency, North-West Frontier Province and Assam. This high degree of spatial variation of cases was largely consistent between years (electronic supplementary material, figure S1). That is, provinces which had experienced large outbreaks in previous years reported a high number of cases in subsequent years. To robustly demonstrate these trends, wavelet analysis was performed on the monthly reported cases within each province. By comparing the annual components of each wavelet, the timing of each outbreak during the year was calculated. This analysis confirmed the annual frequency of outbreaks within each province, and also showed that the timing of these outbreaks was largely consistent over time (electronic supplementary material, figures S2–S4). However, outbreak timing varied considerably between provinces (figure 3). Outbreaks started in the south of the Indian Peninsula in October each year and radiated out towards the north over a period of six to seven months. Simultaneously, annual plague outbreaks cycled between the Burmese native Shan state in September, to Upper Burma in January, finishing in Lower Burma in April. We also found that the timing of outbreaks between most dense population centres (Bangalore Civil and Military Station, Madras City and Calcutta City) and their respective provinces was within one month of one another. By contrast, the timing of outbreaks between Bombay City (Mumbai) and Bombay Presidency differed by approximately six months (electronic supplementary material, figure S5). Figure 3. Timing of plague outbreaks in British India from 1898 to 1949. The mean timing of outbreaks in each province of British India from 1898 to 1949. Outbreaks began in the south of the Indian Peninsula and cascaded out towards the north. At the same time, plague outbreaks cycled around the Burmese Shan state and Burma. The timing of outbreaks was calculated by comparing an annual sinusoidal function with the annual component of plague incidence determined from wavelet decomposition. Grey provinces indicate insufficient plague data to calculate outbreak timing. (Online version in colour.)
We tested the probability of annual outbreaks occurring based on observed climate values for each province, using different thresholds to define an outbreak. A range of outbreak thresholds was tested, from more than 0 deaths per 1 000 000 individuals to 100 deaths per 1 000 000 individuals within a calendar year. For all thresholds tested (B > 1000), moderate mean annual humidity levels of between 60% and 80% were associated with plague outbreaks (figure 4). Outbreaks were 1.9 (95% credible interval (CI) = [1.5, 2.6]) and 2.2 (95% CI = [1.7, 2.8]) times more likely to occur given moderate annual humidity levels (60–80%) compared to lower (40–60%) or higher (80–100%) humidity. Moderate humidity levels were also associated with outbreaks of greater magnitude, although there was substantial variation in outbreak magnitude across all observed mean annual humidity values (electronic supplementary material, figure S6). Outbreaks were also more likely to occur at extreme temperatures and moderately low precipitation (B > 1000), however, this relationship was generally not consistent across different outbreak thresholds (electronic supplementary material, figures S7–S8). From herein, outbreaks were defined as more than 10 per 1 000 000 plague-related deaths in a single year. Figure 4. Humidity effects on outbreak occurrence. Outbreaks had a higher probability of occurring at moderate levels of mean annual humidity. At low and high levels of humidity, outbreaks of any size rarely occurred, B > 1000. Credible intervals were calculated from fitting the mean humidity per year per province to whether outbreaks occurred or not under different thresholds for defining an outbreak. Circles denote the binary data that was fitted, and crosses show the mean probability of an outbreak occurring at binned humidity values. A quadratic polynomial was fitted within a Bayesian framework with weakly informative Cauchy priors on coefficients and assumed binomial error structure using the rstanarm package in R. (Online version in colour.)
We investigated if the systematic pattern (south to north) in the timing of plague outbreaks could be explained by annual oscillations in climate. To that end, temperature, rainfall and humidity data for each province during 1907–1936 was obtained from written records (electronic supplementary material, figure S9 shows these data over time in five different provinces throughout British India). The annual periodicity of the climate data was confirmed using wavelet analysis, and wavelets for each climate variable were compared against annual wavelet components generated from the plague outbreak data. This allowed us to quantify the lag in time between oscillations in climate and plague outbreaks for each year and location. We found that the time delays between oscillations in climate and plague outbreaks were largely consistent across time within each province. That is, within the same region, there was little variation in the lag between outbreaks (when they occurred) and temperature, rainfall and humidity (electronic supplementary material, figure S10). However, these time lags were different between locations and exhibited a similar systematic spatial pattern to the plague outbreak data: shorter time lags were found in the southern tip of the Indian Peninsula and lengthened towards the north of India and modern-day Pakistan. In general, over half of all plague outbreaks lagged four to eight months behind peak rainfall and six to nine months behind peak temperature. The time lags between oscillations in humidity and plague outbreaks were the shortest and most similar across space with delays of three to five months between the peak of humidity and the peak of plague deaths (electronic supplementary material, figure S11). Owing to the variation in lag times between provinces, we investigated if the timing of seasonal oscillations in climate could reliably infer outbreak timing. Across all of British India and all years, oscillations in temperature and rainfall had strong relationships with plague outbreaks (Spearman ρS = 0.75, Pearson ρP = 0.64, B > 1000 and ρS = −0.62, ρP = −0.57, B > 1000, respectively) (electronic supplementary material, figure S12). A simpler linear relationship was detected (ρS = 0.53, ρP = 0.53, B > 1000) between seasonal changes in humidity and the timing of outbreaks (figure 5). At the national scale, outbreaks would occur approximately one month (95% CI = [0.9, 1.2]) later on average throughout British India for every month peak humidity was delayed. Figure 5. Oscillations in humidity were correlated with outbreaks. The timing of seasonal changes in humidity had a strong linear relationship with the timing of plague outbreaks, B > 1000. Later peaks in humidity were associated with later occurrence of plague outbreaks at the national level. Outbreaks were defined as more than 10 plague-related deaths per one million individuals within a single year and province. (Online version in colour.)
Focusing on the variation within a province, we again find that there is a linear relationship between the timing of outbreaks and humidity oscillations (ρS = 0.43, ρP = 0.42, B > 1000) but no evidence to support a relationship with temperature or rainfall (B < 0.001 and B < 0.001, respectively). In particular, within a province we find that a one-month delay to oscillations in humidity translated to a three and a half-week delay in the timing of outbreaks on average (95% CI = [2.5, 4.5]), although there was considerable variability—from two weeks in Madras Presidency to 12 weeks in Burma (electronic supplementary material, figures S13–S15). Zoonotic pathways of Y. pestis transmission are affected by climate, driving plague epidemiology. Optimal environmental factors, such as temperature, precipitation and humidity, can create the ideal conditions from which plague can emerge from zoonotic rodent reservoirs. The exact relationship between climate and plague outbreaks varies between different spatio-temporal contexts and is typically inferred from short time periods. Therefore, in order to understand these relationships in more detail, analyses of longitudinal data across spatial foci are important. Given the global extent of the third plague pandemic [31,56], historical datasets from this period allow us to rigorously explore these effects on large spatio-temporal scales. During the course of the third plague pandemic, India reported over 13 million plague-associated deaths [35,57]. British India was, therefore, an appropriate setting for studying the effects of climate on plague epidemiology. To that end, we were interested in whether climate could explain the occurrence, severity and timing of outbreaks in British India during the third plague pandemic. We started by analysing the timing of plague outbreaks in British India from 1898 to 1949. Similar to the subset of data presented by Yu & Christakos [58], outbreaks would begin annually in the southern tip of the Indian Peninsula and spread up towards the north over six months. Outbreaks would simultaneously start in the east of Burma, and move towards the south. These two separate cycles of plague spread have previously been demonstrated through phylocartographic analysis of Y. pestis isolates from the third pandemic, where strains in Burma were found to be distinct from the those of the Indian Peninsula [31]. Their study inferred different directions of plague spread, however. That is, the data presented in this study implied that Burmese plague originated in the north and spread southward, whereas they found that strains were introduced from Vietnam and spread north towards China. Plague was also introduced into Bombay and Calcutta City via shipping routes during the late nineteenth century [59,60], but annual outbreaks would not start in these regions. These contrasting findings may imply that once plague has been introduced and established among zoonotic reservoirs, the re-emergence of outbreaks is dictated by local environmental and socio-demographic factors. Indeed, Rattus rattus (formerly Mus rattus) and Rattus norvegicus (formerly Mus decumanus) from several regions of British India tested positive for plague in the early twentieth century [61]. However, if local conditions are insufficient to allow the establishment of zoonotic reservoirs, then large scale outbreaks might rarely occur. This was the case in Assam in the north east of India, where exceptionally wet climates throughout the year and poor transport links with neighbouring provinces [62,63], crucial in the spread of plague throughout pre-industrial Europe [64,65], hindered transmission. Outbreaks of plague were more likely to occur at moderate relative humidity levels of between 60% and 80% than at higher or lower humidity, which was independent of how outbreaks were defined in terms of reported deaths. This was in agreement with a similar study looking at the effects of climate during the third pandemic in China [32,66], finding that plague spread fastest at moderate levels of wetness. These findings may be down to the sensitivity of flea egg and larvae survival rates to changes in soil moisture, a proxy for humidity. This could explain the substantial difference in average outbreak timing between Bombay City (Mumbai) and Bombay Presidency. Despite the absence of Bombay City in the presented climate data, Bombay City did indeed have generally higher humidity than the rest of the region encapsulating Bombay Presidency [67]. In contrast to the rest of the country, Bombay City may have, therefore, only been climatically suitable for plague transmission once humidity dropped, offsetting outbreak timing by roughly six months. This feature could be an important consideration when investigating the effects of climate on modern outbreaks, as regions such as Madagascar and the Democratic Republic of Congo have very high humidity throughout the year. The outbreaks in Madagascar in fact do correlate with time periods where humidity subsides, and the suitability of local flea species improves drastically [68]. The timing of plague outbreaks in British India was associated with seasonal changes in humidity. We found that a one month delay in humidity led to an approximate one month time delay in the plague outbreak on average, but with substantial variation between provinces. In addition to driving change in flea suitability, oscillations in humidity may indirectly influence rodent population dynamics via timing of harvest. Harvest has been suggested to draw rodents towards rural agriculture [69], putting farm workers at increased risk of infection. In the case of British India, this may explain the bi-annual nature of plague outbreaks within some regions, such as Burma, where harvest used to occur twice a year [70]. Agricultural land-use has indeed previously been correlated with higher seroprevalence of Y. pestis among rodents compared with other land types [71], but it is unclear where they were infected. That is, either rodents became infected before migrating to agricultural land or were infected from fleas within the fields themselves. Regardless, once harvest is complete, food trade into urban centres could drive infected rat populations into urban areas, sparking a large outbreak. Given our findings on the effects of humidity on the occurrence and timing of plague outbreaks, one might expect there to be equally strong associations between temperature, rainfall and plague. We did not find this in our data. This was in contrast to a study by Xu et al. [27], demonstrating high speeds of plague spread globally during the third plague pandemic influenced by temperature. Our findings may be attributable to temperature and rainfall gradients being fairly flat across most of the Indian Peninsula (in terms of magnitude and timing), so it was unlikely that temperature or rainfall played decisive roles in dictating plague outbreaks in British India during the third pandemic. On the other hand, there was evidence to suggest that rainfall was positively associated with the timing of outbreaks in Burma. This lack of consistent and uniform drivers in our data was unsurprising given previous claims that plague outbreaks can occur under a diverse set of landscapes for a wide range of environmental conditions [72]. The statistical analysis of epidemiological data is fraught with potential limitations. Most notability data quality may be highly variable in quality with discrepancies in both space and time [73]. This is probably an issue with our data, where more populated regions may have had greater resources for observing and recording plague cases. Cases may also have been over-reported where deaths were misdiagnosed as being owing to plague. This is of particular concern in the case of modern-day pneumonic plague outbreaks which can be misdiagnosed in conjunction with other respiratory infections [74]. Pneumonic plague transmission could add additional layers of complexity in trying to understand the drivers of plague epidemiology, as once an outbreak is established, both rodent and flea populations are no longer required. However, the scale of under- or over-reporting over both time and space should not heavily influence our findings, which are primarily concerned with the occurrence and seasonal timing of outbreaks. We stress that climate alone is not enough to infer the size of an outbreak and we also restricted our analyses to relatively large scale outbreaks, eliminating concern about including introduction events which may bias results. Historical records were only available at the provincial level for the monthly plague deaths and climate data. The space–time aggregation of these data across a wide geographical and temporal scale may lead to inappropriate generalizations of climate effects at lower resolutions. That is, the mechanisms that underlie plague epidemiology are complex and poorly understood, such that other abiotic factors may dictate the timing of plague outbreaks over finer spatial scales, and thus the statistical approach taken here may only be applicable within spatial contexts similar to our data. This does raise an area for future work where the influence of climate and socio-ecological factors, such as rat population dynamics, on plague epidemiology should be studied at the sub-provincial level. Altitude and the cold, dry season have indeed been correlated with plague foci in Madagascar [2,75]. Given its climatic diversity [76], perhaps temperature, precipitation and humidity could explain the spatially heterogeneous patterns of plague found in Madagascar [77]. However, without a clear mechanistic understanding linking climate factors to plague epidemiology, the direct application of our results to modern outbreak settings might be inappropriate. In summary, we have shown that humidity was the most important climate factor in dictating the occurrence and timing of plague outbreaks during the third plague pandemic in British India. Humidity data could therefore go a long way in assessing disease risk using spatio-temporal prediction models from recent bubonic plague outbreaks [78]. However, it would also be important to understand how human social contact networks influence plague spread, in order to minimize the impact of pneumonic plague outbreaks [79]. By monitoring local climate factors, our findings could enhance the identification of regions with imminent outbreak risk, thus improving the management of chemoprophylaxis and highlighting settings in which imminent candidate vaccines could be effectively trialled. All data and code are made available through the GitHub repository: wtennant/plague_india [80]. All authors contributed to the writing and revision of the manuscript. We declare we have no competing interests. This research is funded by the Department of Health and Social Care using UK Aid funding and is managed by the National Institute for Health Research (NIHR) (Global Health Research (PR-OD-1017-20002) to M.J.K., M.J.T. and S.E.F.S.). The views expressed in this publication are those of the authors and not necessarily those of the Department of Health and Social Care. S.E.F.S. also greatly acknowledges support from the Medical Research Council (MR/P026400/1). FootnotesElectronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.5003585. © 2020 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original author and source are credited. References
Page 3Earth is urbanizing rapidly, and this results in human-influenced environmental changes that can modify the conditions for life, and the patterns of biodiversity in urban as opposed to rural (non-urban) areas. Urbanization is an ongoing process and a driver of ecosystem change, and cities currently harbour more than half of the global human population [1,2]. Compared with rural areas, urban environments are characterized by elevated levels of air pollution, light pollution (likely of considerable relevance for nocturnal insects such as moths), different visual backgrounds (of relevance for protective coloration, whether via camouflage or warning signals), higher temperatures, fragmented and often monotonous green areas (such as lawns), and depauperate communities with strikingly different species compositions of plants and animals [3–5]. Thus, urbanization provides opportunities to study how species and communities respond to a wide range of environmental changes and rising human pressures [6,7]. Ultimately, this can inform about community assembly rules and the relative importance of stochastic events and deterministic processes in structuring species compositions [8–11]. Although plant species richness tends to increase in cities (mainly due to ornamental flowers in parks and gardens), many animal taxa seem to respond in the opposite way, and the processes underlying the patterns of biodiversity in cities are poorly understood [4,5,12]. Besides potentially influencing species richness, the environmental conditions that characterize urban environments may affect species composition in urban communities. Previous studies indicate that species in urbanized environments do not constitute a random sample of those available in surrounding rural areas [13–15]. This is likely because urban environments impose selection and constitute an ecological filtering process whereby species with certain traits might be favoured and better able to colonize and persist, whereas species with other traits are less likely to colonize and more likely to disappear [7,13,16,17]. More specifically, it can be hypothesized that urbanization imposes spatial sorting that is predictable, in that it favours species with high dispersal enhancing phenotypes, behavioural flexibility, bold and explorative personalities, broad niches, high intra-specific genetic and phenotypic diversity, and generalist lifestyles that buffer against environmental change, promote establishment success, and reduce extinction risk [7,13,16,18–22]. In agreement with the above reasoning, some previous studies suggest that large species and generalists are overrepresented in urban environments compared with small and specialized species [3,7,14,17]. Another recent study reports that species with a high dispersal capacity (inferred from the functionality of the wings), and preference for high temperatures are disproportionately common in cities [23]. These findings are consistent with the notion that the contrasting environmental conditions in urban as opposed to rural areas impose selection and ecological filtering that drives intraspecific evolutionary modifications and induces shifts in community species composition. However, for some traits the consequences of urbanization generate conflicting predictions. For example, the increased ambient temperatures in cities result in increased metabolic costs [24], particularly for larger species, and this is expected to drive shifts towards smaller body sizes. On the other hand, the higher dispersal capacity that comes with increased body size might enable larger individuals and species to better use the scarce and patchily distributed resources in cities [13,14,23]. To resolve this paradox, it is necessary to identify which specific species traits are key to success in urban environments. To determine the direction(s) of the filtering process on different traits (e.g. whether urbanization favours larger or smaller body size) also requires comparisons based on large numbers of species that differ in many attributes. Moths have several features that make them suitable for investigations on this topic. They constitute a species rich and highly diverse group that contribute to ecosystem services and functional diversity by being important in food webs and as pollinators [25–27]. It has even been argued that species composition of moths can be used to assess both the functioning and health of urban ecosystems [28,29]. A number of studies that explore species-specific responses to urbanization among invertebrates across multiple traits have been published mainly from urban areas in central Europe [3,14,17]. Yet, little is known about whether multiple traits have an impact on urban community composition across larger spatial scales, and whether different traits interact to influence urban communities. This may lead to biased conclusions regarding what attributes enable species to cope with environmental change and conditions influenced by human activities, and ultimately hamper projections regarding future biodiversity responses to global change. In this study, we surveyed moth communities in three cities in northern Europe and compared them with neighbouring moth assemblages that comprise species pools of potential colonizers. We analysed occurrence patterns and compared species traits using data for 858 species of moths. Based on predictions from theory and previous empirical findings on how species traits impact colonization and establishment, we expected that communities in urban environments were dominated by multi-dimensional generalist moth species characterized by large distribution ranges, varied habitat preferences, broad diets, variable colour patterns and a long reproductive season [3,7,13,15,16,18–22]. Because urban environments tend to be warm with patchy resource distributions, we also expected species in the cities to have a high thermal tolerance, and either a small (low energy demands) or large (high dispersal capacity) body size [13,14,17,23]. We had no a priori prediction regarding what overwintering stage(s) might be overrepresented in urban environments. We studied 858 night active macro moths from the 14 families Brahmaeidae (1), Cossidae (4), Drepanidae (16), Endromidae (1), Erebidae (86), Geometridae (304), Hepialidae (5), Lasiocampidae (16), Limacodidae (2), Noctuidae (358), Nolidae (11), Notodontidae (36), Saturniidae (2) and Sphingidae (16) (see electronic supplementary material, table S1) in three cities and regions in Europe; Halle (Germany, Sachsen-Anhalt), Kalmar (Sweden, county of Kalmar) and Lund (Sweden, Skåne) (figure 1). The three cities varied somewhat in climate (mean annual temperature and precipitation) and with regard to light pollution, number of inhabitants, availability of green space and population size (table 1). To sample moths in the cities, we used one automatic light trap [32] equipped with a mercury vapour lamp that was in operation from April 2011 to October 2012 in each city. The traps were checked every 10–21 days throughout the season, and all macro moth species were determined to species level and registered. With this method, all species of macro moths caught by our light sources during 3570 light-hours (distributed over all dark hours during the year) were included in our dataset. The traps had a similar position in each city between the city centre and the outskirts with the exact trap locations; Halle: Viktor-Scheffel-Str. 8, 51°29′36.04″ N, 11°58′28.13″ E, Lund: Sölvegatan 37, 55°42′51.21″ N, 13°12′26.27″ E and Kalmar: Landgången 4, 56°39′31.24″ N, 16°21′45.97″ E. Figure 1. Examples of species with different frequency of occurrence in the three cities. From top to bottom, (a) Cosmia trapezina occurred in all three cities, (b) Eilema complana in two, (c) Malacosoma neustria in one and (d) Naenia typica was present in all three surrounding species pools of rural areas but did not occur in any of the studied cities. The three cities in northern Europe included in the present study (indicated by dots), and the three surrounding species pool of rural areas as delimited by dashed lines (e). Photos by Vladimir Kononenko. Map source from Esri. ‘World Topographic Map’, 29 September 2019. (Online version in colour.)
To evaluate the consequences of ecological filtering, we compared the community composition and species traits of moths collected in the three urban environments with the moth assemblages of potential colonizers, as based on previous observations in the corresponding regions. The species composition of the neighbouring macro moth communities (858 species) was determined based on previously published records for the region/province to which each city belonged (electronic supplementary material, table S1). For Halle we used records for the province Sachsen-Anhalt [33], for Lund the province Skåne, and for Kalmar the county of Kalmar [34] was used. The data on the three regional species pools were all based on data from 1990 to 2011. The cities had a rather central position, with the exception of Kalmar that is situated on the coast, facing the Baltic sea in the east—an area that is unlikely to influence the composition of the species pool. Province records have a long lasting tradition, are continuously kept up to date and provide the most reliable information on species occurring in the surroundings of the surveyed cities [33,34]. Many moth species can disperse over long distances (greater than 10 km), and it was assumed that species previously recorded in each province constituted potential colonizers of the respective city [35–38]. However, species that had not been recorded in the provinces in recent years (after 2010), and species known from only one relict population in the province, were excluded from the analyses, as they could not be expected to occur in the studied cities. We extracted data on eight species traits (temperature preference, length of reproductive season, range size, body size, habitat use, dietary breadth, colour pattern variation and overwintering life stage) from the literature for all species studied (see electronic supplementary material, table S1). We used the most recent taxonomy ([39], see electronic supplementary material, table S1). Given that we studied eight predictor traits, it was important to evaluate redundancy among these variables to avoid problems potentially associated with correlated predictors [40]. Redundancy among predictors can exist in two forms: pairwise correlation and multi-collinearity. While correlation measures the degree of association between pairs of variables, multi-collinearity refers to the situation when there is a concurrent relationship between multiple variables (i.e. when some variables can be predicted from a combination of other variables). We pre-processed the set of predictor variables by applying correlation and multi-collinearity analyses. Results did not point to any strong signatures of correlated predictors or multi-collinearity (electronic supplementary material, table S2, figure S1), and we therefore included all eight traits in the remaining statistical analyses [41]. We tested whether the moth species that were caught in the three cities could be characterized by any of the eight species traits or whether urban moth communities comprised species with random trait values. A species was denoted ‘present’ in a city if it had been caught in the city by our traps, and denoted ‘absent’ if it had not been caught in the city but was present in the regional species pool. We used a generalized linear mixed (GLMM) effects model with a binomial error distribution, a log-link function and treatment contrasts. We examined whether results and conclusions were robust or sensitive to alternative analytical approaches by analysing our data in three different ways. In all three analytical approaches, city occupancy (present or absent) was considered the dependent (response) variable, each species recorded in the region/province surrounding the city contributed with one observation, the species traits represented eight explanatory variables, and the three cities and their associated provinces were considered as separate samples (replicates). The sample sizes (number of species recorded in each region) were 853 for Halle, 718 for Lund and 695 for Kalmar. In the first analysis, all traits were combined and analysed together in the same GLMM. In this approach, city occupancy (0/1) was treated as a binary response variable, and temperature preference, length of reproductive season, range size, body size, habitat use, dietary breadth, colour pattern variation and overwintering life stage were treated as fixed explanatory variables (see above). Region was included as a fixed explanatory factor to account for variation in community composition among the three cities and regions. We did not evaluate the contributions of two-way and higher-order interactions between the explanatory traits, to avoid problems associated with over-parametrizing statistical models. Species nested in region and species nested in taxonomic family were included as random terms in the model to partially account for greater similarity in the response variable among species that are more similar in general ecology and life history because they are more closely related [42–45]. The models were built based on AIC (Akaike's information criterion), by comparing all possible models with different combinations of the explanatory variables. We present the model with the lowest AIC as it is considered the model that best fit our data. We used the glmer function in the lme4 package for the GLMMs and the Anova function in the car package to test each main effect after the other main effects (type II) in the GLMMs. We used the function r.squaredGLMM described in Nakagawa & Schielzeth [46] to obtain values for the explained variance in the final GLMMs. Second, we evaluated whether results and conclusions regarding the roles of the eight species traits were similar or varied among the different cities. To that end, we performed a separate GLMM for each city/region. In these analyses, all eight species traits were included as fixed explanatory variables, taxonomic family was included as a random term and model selection was performed as described above. Third, we evaluated whether and how results and conclusions regarding city occupancy changed depending on whether species traits were analysed together in one model or one at a time in separate models. For that purpose, we performed eight separate GLMs, one for each trait. City was included as a fixed factor to account for variation in community composition among the three cities and regions. Because we analysed each trait separately, significant associations may occur by chance. Adjusting critical significance levels for multiple tests has problems [47,48]. Nevertheless, we indicate in the results sections whether the reported test results remained statistically significant after sequential Bonferroni corrections, thereby enabling readers to judge for themselves. Besides the approaches outlined above that were used to evaluate the statistical significance of associations of species traits with city occupancy, we used an additional, separate statistical method based on ordination to generate plots that enabled visualization of results. To illustrate and describe the trait composition that characterizes moth species collected in urban environments and how they differ from moth assemblages in neighbouring areas we used non-metric multidimensional scaling (NMDS) [49]. For this, we used a species by trait matrix that included presence/absence and the number of cities occupied by each species. The NMDS was performed with the R package vegan [50] using the Bray–Curtis dissimilarity measure. Of the 858 moth species recorded in the combined species pool, 392 species were caught in at least one of the cities, 196 species were caught in one city only and 90 species were caught in all three cities. In Halle, we caught 286 of the 853 species that were included in the species pool, in Lund 215 of the 718, species in the species pool, and in Kalmar we caught 178 of the 695 species (electronic supplementary material, table S1). When data for all three cities and all eight trait variables were analysed together, the final best fitting model included all traits except body size (AIC = 2276, table 2), and provided a better representation of the data compared to the null model that only included the random terms (AIC = 2661.6). Results thus provide strong evidence for urban filtering across multiple species traits. Overall, species in the cities had significantly larger distribution ranges, more variable colour patterns, a longer reproductive season, higher temperature preferences, broader diets and were more inclined to be habitat generalists (table 2, figure 2). Further, species in the cities were more likely to overwinter as an egg compared to imago, larvae and pupae (table 2, figure 2). Species recorded in the surrounding regions were also more likely to occur in the city of Halle, compared to both Lund and Kalmar (table 2, figure 2). Figure 2. Odds ratios (dots ± 95% CI denoted by vertical lines) for the relationship between moths present in the cities (compared with a species pool of surrounding rural areas) and the seven species traits that remained in the model with the lowest AIC value (table 2). The vertical line shows the odds ratio of 1.0. Each of the categorical variables are compared to a reference category i.e. habitat (open, forest species to generalist species), overwintering (larvae, pupae, imago to egg) and Lund/Kalmar to Halle.
The results and conclusions regarding the roles of the explanatory species traits for city occupancy were largely robust to choice of analytical approach (tables 2–4). Although statistically important variables, coefficient estimates and parameter values varied somewhat between the statistical models, the results were surprisingly similar. However, the results for body size and temperature preference varied depending on whether they were analysed together with other traits in the same model or analysed alone (table 4). The association of city occupancy with overwintering stage did not remain statistically significant when adjusting for multiple comparisons (table 4). Comparisons of mean body sizes showed that, when the contributions of all other traits were not taken into consideration, species occurring in the city were larger on average compared with species in the region (table 4). However, this effect of body size did not manifest when all traits were analysed together (tables 2 and 3). The conclusion was opposite for temperature preferences. When all traits were analysed together in the multivariate GLMM, the results indicated that moth species caught in the city were more likely to be thermophilic (tables 2 and 3), but temperature preference was not significant in the univariate GLMs (table 4). The results for the remaining six traits were qualitatively similar (had the same direction) when analysed separately and together with the other traits (tables 2–4).
The results generated by the NMDS illustrated and confirmed that urban moth communities were structured based on species traits (electronic supplementary material, figure S2). The results further indicated that urban communities differed from the regional species pool. This was evident for each of the three cities/regions and also for the number of cities the species occupied (from zero to three, see electronic supplementary material, figure S2). We explored whether moth communities in urban areas primarily included generalist species with broad niches that are able to cope with more novel, variable, fragmented, warmer and unpredictable environments shaped by human activities. Results showed that urban moth communities consisted largely of multi-dimensional generalists with larger distribution ranges, more variable colour patterns, longer reproductive seasons, broader diets, tended to occupy more habitat types, were more likely to overwinter as an egg, and were also more thermophilic compared with species of moths recorded in surrounding areas. From a long-term eco-evolutionary perspective, cities may be thought of as offering novel habitats and resources ready to be colonized and used by species that have the capacity to do so [51,52]. On average, only about one third of the moth species that made up the pool of potential colonizers available in surrounding rural areas were present (caught) in the three cities included in our study. Urban environments thus seem to impose a filtering process that negatively influences species richness. This finding is in agreement with studies of butterfly and bird communities [15,53], and moth communities that were reduced by 82% in urban areas compared to rural areas in Belgium [17]. Another key finding that emerged from our analyses is that moth communities in the three cities studied here did not constitute random subsamples of the species that made up the pool of potential colonizers recorded in the surrounding rural areas. Instead, results from the comparisons of species characteristics between city exploiters and moth species in rural areas pointed to an important role of spatial sorting and species filtering, thus suggesting an important role of deterministic processes rather than stochastic events for community assembly [8–11]. Specifically, our results suggest that urbanization imposes a strong filtering process in favour of multi-dimensional generalist species characterized by large range size, habitat and diet generalists, and a high intraspecific diversity (table 4; electronic supplementary material, table S3 and figure S2). As a result of continued urbanization and biotic homogenization [17], future communities are expected to comprise fewer species and an overrepresentation of species having broader niches and more generalized lifestyles, and a lower incidence of specialized species. This is likely to apply across spatial scales, taxonomic groups and trophic levels, based on the finding that results seem to be general among birds [54] and several groups of invertebrates [17,23]. For range size, length of reproductive season and habitat use, the results regarding how different phenotypic traits responded to the urbanization filtering process were qualitatively very similar across the three cities. Body size was only important in one city. For the other traits the model fit improved in two out of the three cities (table 3). The reasons for these minor discrepancies are probably mainly due to the lower statistical power when data are analysed separately for each city. Further, the discrepancies they might reflect differences in human population size, area, proportion of green space, annual average temperature, precipitation, and light pollution. Nevertheless, that results were comparable for most traits across the three cities (table 4; electronic supplementary material, table S3), suggests that the community assembly rules [8–11] that influenced species composition of moths were similar, and further indicates that deterministic ecological and environmental filtering processes were more important than stochastic events. This adds an important layer of generalization to the issue under investigation, and suggests that the effects that the fundamental ecological processes involved (e.g. dispersal, colonization and extinction) have on species with different characteristics are predictable, at least in the sense that they are repeatable in space. Species traits have also been found to be reliable predictors of spatial trends and temporal dynamics across several taxonomic groups [3,38,55–59]. The smallest city (in terms of both human population size and area) had the lowest species number and the lowest proportion of the species pool present (and vice versa for the results for the largest city). This is a surprising finding as the opposite could be expected if the regional species pool represents the source of potential colonizers, and if urbanization can be seen as a filtering process per se. The equilibrium theory of island biogeography [60] posits that species richness should be higher in larger cities. That species richness increased with city size might reflect in part that more habitats are available in larger cities, but independent of the city the species in the cities are characterized by multidimensional generalists. That the number of species per urban square kilometre was highest in the smallest city might reflect in part that trapping effort per city area decreased with increasing city size (table 1). A reassuring finding (from a methodological perspective) was that results and conclusions regarding the associations of different species traits with city occupancy were largely independent of whether the traits were analysed one at a time or together, except for body size and temperature preference. This suggests that variation among species in at least six of the eight analysed traits contribute in important ways to the ability of moth species to colonize, cope with and persist in challenging urban environments. That species with large geographical range distributions were overrepresented in the cities was expected, as they occur across larger regions, including urban areas and cities, compared to species with restricted ranges [17,61]. That species with a high colour pattern variation were overrepresented in cities is consistent with previous evidence that more variable colour patterns in moths is associated with larger intraspecific trait variation also in other (morphological, behavioural, and life history) phenotypic dimensions [62], and with increased colonization success, more stable population dynamics, and decreased extinction risk [42,43,45]. Moreover, light pollution [63] and non-native plant species [5] might create more heterogeneous and novel visual backgrounds in cities that are more suitable for species with variable colour patterns [64]. Altered predator communities in cities have been found in a previous study [23], and this might favour species with variable colour patterns. A longer reproductive season may be favourable in cities as it extends the possibilities to use and find resources over time and space, compared to species with a short reproductive season. That diet and habitat generalists were overrepresented in the cities may be reflective of that food and habitat resources are novel, diverse, and only temporarily available [5,12,17,52]. Species associated to open habitats were less likely to be present in the cities (odds ratio 0.37 to generalist species), probably reflecting that most open habitats in cities are strictly managed lawns and parks with low overall species richness of plants [51]. In contrast to our results, a recent study reports that open habitat species are overrepresented in urban areas, and speculate that this could be an effect of the increased heat tolerance among open habitat species [17]. In view of these last findings, a change in cities towards less intensively managed areas may allow for a richer biodiversity [12,51,65]. It remains to be explored whether the declining species richness and biotic homogenization observed in this study allows for accurate predictions regarding dominating species traits in urban moth communities in other areas, on other continents, and in urban communities of other types of organisms. Our conclusion regarding the role of body size differed depending on analytical approach. Several previous studies report that urbanization favours larger species, possibly because larger size is associated with a superior dispersal and greater ability to use fragmented habitats and patchily distributed resources [3,7,14,23]. For example, Merckx et al. [13] report on community-level shifts in body size (forewing length) based on data for 23 species of butterflies and 202 species of macro-moths (wing span). Similarly, other studies report on both community-level and intra-specific shifts towards larger species and individuals in urban communities, based on individual measures and community weighted wing length data [14,17]. Community weighted variables used in previous studies might potentially suffer from the drawback that it is difficult to combine multiple community traits and taxonomic levels in the same statistical model [13,66]. In agreement with this conjecture, we found that when body size was analysed alone, moth species that occurred in the city were larger on average compared with species in the surrounding rural areas. However, when the role of body size was evaluated together with the other seven traits, the city, and taxonomic family, the association between city occupancy and large body size disappeared. This may be reflective of that larger size, while perhaps not being very important in itself, may be an adequate proxy for ecological generalization, and for taxonomic groups and traits that are well suited for life in urban environments. For temperature preferences the pattern was opposite, an association of city occupancy with thermophily was indicated by the multivariate but not by the univariate analyses. This emphasizes the importance of using both univariate and multivariate analytical approaches in ecological studies to better understand the importance of different predictor traits on species occurrence patterns. The depauperate urban moth communities, with only about one-third of the species in rural areas occurring also in urban environments, is an alarming finding. Considering that moths contribute with many important ecosystem services [25–29], this points to a severe future biodiversity decline and biotic homogenization that may have far reaching implications for the functioning of communities and ecosystems. We found that urban moth communities were not random subsamples of neighbouring communities. Instead, our results point to the conclusion that urbanization imposes a spatial sorting filtering process that reduces species richness and results in biotic homogenization, favouring thermophilic and multi-dimensional generalist species characterized by high intraspecific diversity. The associations between species traits and city occupancy were mostly qualitatively robust and repeatable across the three regions (albeit with some exceptions). This is indicative of generality and predictability of the ecological and evolutionary drivers involved in the community assembly process. Future investigations of cities across other latitudes and with higher population densities will reveal whether our results are more broadly applicable, and whether the mechanistic drivers that structure urban communities may continue to inform about eco-evolutionary consequences and biodiversity responses to global environmental change. It can be hypothesized that species possessing those trait value combinations that enable them to cope with life in the city will also be particularly successful in the future when urbanization continues. Data are provided in the electronic supplementary material. M.F. and L.B.P. conceived the study. M.F., L.B.P. and P.-E.B. collected data from the field and information on species traits, M.F. analysed the data. A.F. and M.F. wrote the first draft. All authors contributed to the final version and approved the submitted manuscript. We declare we have no competing interests. This work was supported by Linnaeus University (grants to A.F., P.-E.B. and M.F.), the Swedish Research Council Formas (grant number 2018-02846) (to M.F. and A.F.), the Swedish Environmental Protection Agency (to L.B.P.), Birgit and Hellmuth Hertz Foundation (to M.F.) and the Crafoord Foundation (to M.F.). We thank Cordula Vogel, Mikael Molander and Martin Lindner for help in the field and with extracting information of the species, and Thomas Merckx and two anonymous reviewers for comments on the manuscript. Victor Johannsson provided helpful statistical insights and help with the NMDS ordination. FootnotesElectronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.4991894. © 2020 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original author and source are credited. References
Page 4The ability of ecosystems to reliably provide biological products and services to humanity is being threatened by climatic changes and species loss [1–5], arousing interests in biodiversity–stability relationships. As an important ecosystem function, community biomass tracks environmental fluctuations [6–8]. This temporal variability is frequently estimated with the temporal coefficient of variation (CV = σ/μ, the inverse of community stability) [9–13], and, based on theoretical and experimental studies [14–16], is expected to be lower in species-rich than in species-poor communities owing to biotic mechanisms. First, asynchronous species dynamics can stabilize communities because the decreased biomass in one species can be compensated by increases in others [8,17–20]. Second, the mean–variance scaling relationship [21] suggests a stabilized community by distributing the total community biomass among more species when the scaling power ranges from 1 to 2 [10,13,22]. Third, the positive biodiversity–biomass relationship suggests that high species richness can stabilize communities by increasing total biomass, a phenomenon termed the overyielding effect [3,13]. Although these biotic mechanisms have been proposed to stabilize communities in theoretical analyses [10,12,20,23–25] and have been evidenced by single-site observational [8] and experimental [14,15,17,26] studies, as well as meta-analyses [16,21,27,28], they are still argued to be irrelevant in natural communities, especially at large spatial scales [2,29,30]. Community stability can vary spatially, potentially owing to spatial variations in abiotic and biotic factors. For example, a study in North America showed that community stability of grassland was lower than that of desert and forest because of its intermediate precipitation CV and potential productivity [7], suggesting that both abiotic and biotic factors affected community stability in natural ecosystems. Biotic stability mechanisms may be sensitive to climatic variation [27,28,31–34], masking their relevance in natural ecosystems. This environmental dependency suggests that stabilizing effects of biotic mechanisms on communities may vary across sites and ecosystems, and thus stabilizing effects found in a single-site study may not extrapolate to multiple sites across larger spatial scales. In addition, abiotic factors can affect community species composition [17,35] or even outweigh biotic factors in driving species richness [1,27,36] and species dynamics [28,31] at the regional scale, thus further modifying biotic stability mechanisms. Finally, abiotic factors may exert their influence on community stability via biotic factors or, in other words, biotic mechanisms may mediate the effects of abiotic factors on community stability. Therefore, investigating how biotic mechanisms affect community stability under varied environmental conditions and community species richness and composition in a multi-site study may give us deeper insights into biotic stability mechanisms in natural ecosystems than do single-site studies. We conducted a multi-site observational study for five consecutive years across Inner Mongolian grassland in China (total area 78.8 million hectares [37]) (electronic supplementary material, figure S1). The 23 study sites were characterized by large between-site variation in climatic factors (e.g. precipitation and its interannual variation) and biotic factors (e.g. community biomass and species richness and composition) (electronic supplementary material, table S1 and figure S1). The Inner Mongolian grassland is a typical part of the Eurasian grassland biome and crucial in providing biological products and services to human populations living there [33,38]. In this region, precipitation is the dominant driver of biomass [8,33,36,39,40], species richness and species composition [36,41] in communities. Observational studies have shown that precipitation in Inner Mongolian grassland has dramatically changed during the past decades [42,43], while the ecological consequences of precipitation changes are still not fully understood. To investigate the community stability in this region, we employed a novel theoretical model relating its inverse (the community temporal CV) to species synchrony and weighted (by relative species biomass abundance) average species temporal CV, synthesizing the mean–variance scaling effect and the overyielding effect [13] (electronic supplementary material, appendices A.1 and A.2). We analysed whether the community stability and its biotic mechanisms depended on climatic factors (precipitation and its interannual variation) and how community stability was affected by those biotic mechanisms and species richness. We separated the biotic stability mechanisms into different species-abundance groups (electronic supplementary material, appendices A.3 and A.4) and hypothesized that community stability was more strongly affected by the most abundant, i.e. dominant, species than by total plant species richness. This is because abundant species contribute strongly to dynamic processes in communities of high unevenness [13,25,44] and the studied region is characterized by such high unevenness of species biomasses [8,33] (electronic supplementary material, figure S2). The Inner Mongolian temperate grassland has a continental monsoon climate with a short and cool growing season (from May to October) and a long and cold non-growing season (from November to April) [36,45]. During the period of this study (from 2012 to 2016), the mean growing-season precipitation ranged from 186.2 to 398.0 mm, corresponding to about 90% of the annual precipitation (electronic supplementary material, table S1). The Inner Mongolian grassland includes three main vegetation types: meadow steppe, typical steppe and desert steppe (electronic supplementary material, figure S1). The meadow steppe is dominated by perennial grasses such as Stipa baicalensis and Leymus chinensis and perennial forbs such as Convolvulus ammannii. The dominant species of the typical steppe are perennial grasses such as Stipa grandis, Leymus chinensis and Stipa krylovii. The desert steppe is dominated by perennial grasses such as Stipa caucasica and perennial forbs such as Allium polyrhizum (electronic supplementary material, table S1). In this region, aboveground community biomass varied considerably from 20.9 to 180.9 g m–2 and species richness ranged from 6 to 25 (electronic supplementary material, table S1). The lowest community biomass (20.9–43.7 g m−2) and species richness (6–13) occurred in desert steppe and the highest values in meadow steppe (101.6–180.9 g m−2 for biomass and 11–25 for richness) (electronic supplementary material, table S1). In 2012, we established a 23-site observational study along a transect across the Inner Mongolian grassland of China (latitudes ranged from 39.34 to 49.96°N and longitudes from 107.56 to 120.12°E) (electronic supplementary material, table S1 and figure S1). We recorded the geographical location (latitude and longitude) of these sites and resampled them in the following 4 years (2013–2016), resulting in a 5-year-long time series of field observations. For more than half of our sites (13 out of 23 sites), data for the whole 5 years were collected. For some sites (10 out of 23 sites), only data for 3 or 4 years were collected (two sites with data for 3 years because of land-use change and eight sites with data for 4 years because of mowing before the community survey could have been done) (electronic supplementary material, figure S3). A recent 39-site meta-analysis showed that investigating biodiversity–stability relationship with time series of less than or equal to 4 years can produce similar results to those using longer time series [46]. In addition, multiple short time series to some extent may compensate few long time series. We, therefore, expected that our dataset could provide reliable insights into the biodiversity–stability relationship of the studied region. Plant communities were surveyed between late July and early August in each year with a method that has a well-documented efficiency to estimate aboveground biomass and plant species richness in grassland ecosystems [36,40,47]. To appropriately represent the natural community, the surveyed plant community at each site was randomly selected in each year, excluding areas with anthropogenic disturbances, e.g. heavy grazing or mowing. We positioned a plot of 10 × 10 m at each site and surveyed three 1 × 1 m quadrats along the diagonal. Subsequently, all living plant material in each quadrat was harvested and sorted into species. All material was oven-dried and weighed to obtain aboveground biomass and calculate effective species richness based on species biomass abundance (see below). To obtain site-specific precipitation data, we collected the monthly climatic data from 119 climate stations across Inner Mongolia, and then calculated site-specific monthly precipitation using a kriging method with a 2 km resolution digital elevation model in ArcGIS software (Environmental Systems Research Institute, Redlands, CA, USA). A previous study has shown that the data interpolated using this method correlate well with in situ measured climatic data [47]. Using these interpolated site-specific precipitation data, we calculated the mean growing-season precipitation and its interannual variation (estimated as the temporal CV). In this study, we used these growing-season climatic variables because plants were most active during this period. Based on a recent theoretical model, we related the community temporal CV to the species synchrony and the weighted average species temporal CV [13,25] (electronic supplementary material, appendices A.1 and A.2). In the current study, these two terms were estimated with either all species or only dominant species (relative species biomass greater than or equal to 5%), defined as the dominant species synchrony and the weighted average dominant species temporal CV (electronic supplementary material, appendices A.3 and A.4). According to a recent theoretical study, the weighted average species temporal CV can be affected by the mean–variance scaling relationship and the overyielding effect [13]. Here, the mean–variance scaling relationship is defined as the power function between temporal variance and mean biomass [13,21] and overyielding is defined as a positive effect of species richness on biomass [3,13]. The theoretical model shows that these two biotic stability mechanisms can interactively affect the weighted average species temporal CV, thus underpinning the effects of species richness [13]. In addition, the theoretical model suggests a weak effect of species richness on the weighted average species temporal CV when the mean–variance scaling relationship has a coefficient close to 2 [13]. In the current study, species richness was defined as the multi-year average number of species recorded in a 1 m2 quadrat. Considering the high unevenness of species biomasses in the studied grassland communities, we also used a measure of effective species richness, the antilog of the Shannon–Wiener diversity. This measure reflects how many species with an even abundance distribution would produce the same Shannon–Wiener diversity as observed for the actual uneven community [48]. These two methods were also used to estimate the species richness and effective species richness of the dominant species. The community temporal CV was estimated using aboveground biomass over the 5 years of the survey without detrending because biomass had no significant linear temporal trend (assessed using linear regression between biomass and year) except for one of the 23 study sites (electronic supplementary material, table S2). Taylor's power law [21] was used to estimate the mean–variance scaling coefficient (electronic supplementary material, figure S4). We did not explicitly estimate the strength of the overyielding effect [3] as we had no monoculture treatments, but could detect it via positive slopes of linear regressions between biomass as dependent and species richness as the independent variable (electronic supplementary material, figure S5). We calculated the correlation coefficients between the climatic factors (precipitation and its interannual variation), biotic factors (species richness and effective species richness), biotic mechanisms (mean–variance scaling, weighted average species temporal CV and species synchrony) and the community temporal CV to develop causal hypotheses about relations between variables. Individual relationships were plotted and analysed with linear regression to assess how climatic and biotic factors directly influenced biotic stability mechanisms and the community temporal CV, and how biotic stability mechanisms directly influenced the community temporal CV. To combine causal hypotheses about direct and indirect effects, we incorporated them into structural equation models (SEMs) and displayed their results with path-analysis graphs, using the lavaan package [49] of R 3.4.0 [50]. Specifically, we constructed SEMs that deliberately stayed as close as possible to a priori hypotheses proposed to be essential biotic stability mechanisms [13,19,20,27]. We did this even at the cost that the overall model fits might show significant deviations from a saturated model. Model-fit statistics such as χ2-tests or the goodness-of-fit index (GFI) comparing the deviation of a current SEM to a full SEM without residual degrees of freedom were only used as an additional guide, but we avoided searching for a best model post hoc (electronic supplementary material, appendix B). All the above analyses were carried out with all species or only the dominant species included in biodiversity measures and biotic stability mechanisms, i.e. synchronous dynamics and weighted average species temporal CV of dominant species. All statistical analyses were conducted using R 3.4.0 [50] with the graphics package for plotting figures. For the correlation analyses, regression analyses and SEMs, relationships and pathways were considered significant if p < 0.05. The mean growing-season precipitation and its interannual variation were strong drivers of species diversity of the studied Inner Mongolian grassland sites. Specifically, both species richness and effective species richness were positively associated with growing-season precipitation (figure 1a). In addition, the richness and effective richness of dominant species were negatively associated with the interannual variation in growing-season precipitation (figure 1b). Figure 1. Correlation matrix for climatic factors (mean growing-season precipitation and its interannual variation), biodiversity indices (species richness and effective species richness), biotic stability mechanisms (mean-variance scaling exponent, weighted average species temporal coefficient of variation (CV) and species synchrony) and community temporal CV. (a) Correlation matrix for variables calculated with all species and (b) correlation matrix for variables calculated only with dominant species (except climatic factors and community temporal CV). Black numbers with coloured background represent significant (p < 0.05) correlations and grey numbers with white background represent non-significant (p > 0.05) correlations. (Online version in colour.)
The community temporal CV was significantly affected by climatic factors but independent of species diversity. Specifically, growing-season precipitation had a negative effect on the community temporal CV, while its interannual variation had no significant effect (figure 1). In addition, species richness and effective species richness of all and of dominant species did not significantly affect the community temporal CV (figures 1 and 2a–d). Figure 2. Community temporal coefficient of variation (CV) in relation to species richness, effective species richness, weighted average species temporal CV, species synchrony (a,c,e,g, respectively) and the corresponding relations for dominant species only (b,d,f,h). Black solid lines represent significant linear relationships (p < 0.05) and grey dashed lines represent non-significant linear relationships (p > 0.05).
The community temporal CV was positively related to the mean–variance scaling exponent, the weighted average species temporal CV and species synchrony of all and of dominant species (figures 1 and 2e–h). The weighted average species temporal CV of all species was not significantly influenced by growing-season precipitation but if it was calculated only for dominant species a negative relationship was found (figures 1 and 3a,b). Furthermore, the weighted average species temporal CV of all and of only the dominant species was not significantly related to the corresponding measures of species richness (figures 1 and 3c–f). However, the weighted average species temporal CV of all and of only the dominant species was positively associated with the mean–variance scaling exponent (figures 1 and 3g,h). The mean–variance scaling exponent was independent of climatic factors, i.e. the growing-season precipitation and its interannual variation, and species diversity indices measured with all species or only with dominant species (figure 1). We did not explicitly estimate the overyielding effect [3] but found significantly positive species richness–community biomass relationships at nine out of 23 sites (and only at one site was the relationship significantly negative; electronic supplementary material, figure S5), indicating that overyielding effects did occur. Figure 3. Weighted average species temporal coefficient of variation (CV) in relation to precipitation, species richness, effective species richness and mean–variance scaling exponent (a,c,e and g, respectively) and weighted average dominant species temporal CV in relation to precipitation, richness and effective richness of dominant species and mean–variance scaling exponent (b,d,f and h, respectively). Black solid lines represent significant linear relationships (p < 0.05) and grey dashed lines represent non-significant linear relationships (p > 0.05).
Both the community-wide species synchrony and the dominant species synchrony were positively affected by the interannual variation in growing-season precipitation (figures 1 and 4a,b). In addition, the community-wide species synchrony was negatively associated with effective species richness (figures 1a and 4e). The dominant species synchrony was negatively associated with both the dominant species richness and dominant effective species richness (figures 1b and 4d,f). Figure 4. Species synchrony in relation to the interannual variation in precipitation (precipitation CV), species richness and effective species richness (a,c and e, respectively) and dominant species synchrony in relation to precipitation CV and richness and effective richness of dominant species (b,d and f, respectively). Black solid lines represent significant linear relationships (p < 0.05) and grey dashed lines represent non-significant linear relationships (p > 0.05).
To test the hypotheses that the weighted average species temporal CV and species synchrony directly influenced the community temporal CV, we conducted path analysis to relate the community temporal CV to climatic factors, species diversity indices and biotic stability mechanisms using SEMs (electronic supplementary material, appendix B, table S3 and figure S6). This path analysis did confirm the hypotheses and had a total explanatory power of 0.87 (electronic supplementary material, appendix B; figure 5a). In addition, the mean–variance scaling exponent and the interannual variation in growing-season precipitation increased the community temporal CV indirectly via increasing the weighted average species temporal CV and the species synchrony, respectively (electronic supplementary material, appendix B; figure 5a). Effective species richness indirectly reduced the community temporal CV via decreasing species synchrony in this model (figure 5a). When we replaced effective species richness with species richness its effect on species synchrony was no longer significant (figure 5c). Figure 5. Path-analytic representations of structural equation models (SEMs) for relating the community temporal coefficient of variation (CV) to climatic factors (mean growing-season precipitation and its interannual variation), biodiversity indices (species richness and effective species richness) and biotic stability mechanisms (mean-variance scaling exponent, weighted average species temporal CV and species synchrony). (a) The best SEM based on our a priori hypotheses and statistical analysis (electronic supplementary material, appendix B, table S3 and figure S6), and (c) an SEM model in which effective species richness is replaced with uncorrected species richness. (b,d) SEMs corresponding to those in (a,c), respectively, but where biodiversity indices and biotic stability mechanisms were calculated only with dominant species (dominant species richness, the effective richness of dominant species, weighted average dominant species temporal CV and dominant species synchrony). Solid and dashed arrows represent positive and negative paths, respectively. Arrows with black numbers and asterisks represent significant paths (p < 0.05). Arrows with grey numbers represent non-significant paths (n.s., p > 0.05) that were nevertheless included in the SEMs because they corresponded to a priori hypotheses for causal relationships. All arrows are scaled in relation to the strength of the relationship, with numbers showing the standard path coefficients (i.e. indicating by how many standard deviations (s.d.) the variable at the end of an arrow would change if the variable at the beginning of the arrow would be changed by 1 s.d.). R2 values are proportions of variance explained by dependent variables in the model. Model-fit statistics such as χ2-test and GFI are shown in each panel. The significance level of each path is indicated by the number of asterisks (*P < 0.05, **p < 0.01, ***p < 0.001). (Online version in colour.)
Based on the above SEMs, we furtherly analysed the effect of the dominant species alone on the community temporal CV by replacing community-wide biotic stability mechanisms and biodiversity indices with their counterparts calculated only with dominant species. These analyses showed that the weighted average dominant species temporal CV and the dominant species synchrony positively affected the community temporal CV with a total explanatory power of 0.68 (figure 5b,d). The mean–variance scaling exponent increased the community temporal CV via increasing the weighted average dominant species temporal CV; and growing-season precipitation decreased the community temporal CV via reducing the weighted average dominant species temporal CV (figure 5b,d). Furthermore, both the effective richness (figure 5b) and the uncorrected richness of dominant species (figure 5d) decreased the community temporal CV via reducing the synchrony of dominant species. In the present study, we investigated stabilizing effects of biotic mechanisms on natural community biomass at the regional scale of Inner Mongolia and analysed their dependencies on climatic factors, species diversity and dominant species dynamics. We found that the investigated biotic mechanisms strongly affected community stability in the 23 natural grasslands across a gradient of yearly precipitation and its variation. However, in contrast to expectations based on previous studies [27], the latter had negative rather than positive indirect effects on community stability because yearly precipitation fluctuations increased rather than decreased the synchrony of population dynamics of the different species within communities. This could compromise the reliability of the studied Inner Mongolian grassland in providing biological products and services to human populations under the ongoing increasing precipitation variability [42,43], as we discuss further below. A recent theoretical analysis showed that both non-significant and negative biodiversity–stability relationships were possible when species synchrony, mean–variance scaling and overyielding interactively affected community stability [13]. In the present study, both species synchrony and the mean–variance scaling had significant effects on community stability. Furthermore, nine out of our 23 study sites showed significantly positive biodiversity–biomass relationships indicating overyielding (electronic supplementary material, figure S5). Community stability was independent of overall species richness but indirectly positively affected by a higher richness of dominant species, which decreased dominant species synchrony, which in turn decreased the community temporal CV. This indicates that in natural communities with highly uneven abundance distributions, rare species, which might be sink species not able to maintain independent populations in the community, can mask the influence of biodiversity variables on biotic stability mechanisms. Generally, theoretical and empirical studies predicted and found positive effects of species richness on community stability [14,15,27,28,32]. No effects have been reported from a recent single-site study in the same region as that of our study [33], potentially again for the above reason of highly uneven species abundance distributions in these grasslands. Previous theoretical work found that unevenness can indeed weaken biodiversity–stability relationships [9,22], because the most diverse components of a community, namely rare species, may have limited effects on community stability owing to their low abundances [51]. This theoretical prediction is supported by a growing number of experimental investigations, as well as the current study, showing weak or non-significant biodiversity–stability relationships when dominant species regulate community stability [32–34,52–55]. In the present study, the dominant species as a group had low but variable richness across the 23 sites (1–5 species per square metre, accounting for 64.2–96.8% of community biomass; mean 82.8%) (electronic supplementary material, table S1). The weighted average dominant species temporal CV and the dominant species synchrony significantly impacted the community stability, with an explanatory power slightly lower than that of using community-wide counterparts and much higher than that of common- and rare-species groups (electronic supplementary material, figures S7 and S8). This suggests that theoretical studies assuming evenly distributed species abundances [10,12,18,23] may overestimate the regulatory effect of overall species richness on community stability, or, in other terms, that especially in natural ecosystems with highly uneven species abundance distributions it may be more appropriate to base predictions of community stability on the richness and population dynamics of the dominant species, as suggested above. The weighted average species temporal CV provided the most important biotic stability mechanism across the 23 sites of the Inner Mongolian grassland, which was independent of overall or dominant species richness but was positively associated with the mean–variance scaling exponent. The mean–variance scaling has commonly been omitted in previous studies [26,32,33], while our results indicate that this biotic stability mechanism generally exists in natural grassland communities. Indeed, a recent theoretical analysis showed that the mean–variance scaling can determine the sign of the relationship between species richness and the weighted average species temporal CV [13]. The authors of that analysis found that species richness will negatively impact the weighted average species temporal CV when the mean–variance scaling has an exponent ranging from 1 to 2 [56]. However, a positive effect will occur when it is greater than 2 (some studies showed that such high values are not impossible, see e.g. [57,58]). In the present study, the estimated mean–variance scaling exponent had a mean value of 1.72 (ranging from 1.41 to 1.98). This value is consistent with the commonly considered range of 1–2 and the reported value (1.73) in a recent single-site study in this region [34]. However, this value is close to 2, and thus may in part explain the non-significant relationship between the weighted average species temporal CV and species richness. The weighted average species temporal CV was strongly affected by the dominant species group and independent of the most diverse component of the community, the rare-species group (electronic supplementary material, figure S7), which was likely responsible for the lack of significant effects of total species richness. In addition, we found that higher growing-season precipitation can stabilize communities via decreasing the weighted average dominant species temporal CV. This is because of its stronger stimulation of the mean biomass than of its standard deviation (electronic supplementary material, figure S9), therefore, decreasing the standard deviation-to-mean ratio. More importantly, dominant species showed higher stability than other species (electronic supplementary material, figure S10), indicating that they are better able to maintain a stable biomass in a fluctuating environment than other species, potentially owing to their better abilities in acquiring nutrients, water and light via well-developed root systems and taller canopy [17,59,60]. Thus, these findings suggest that in the studied grassland communities dominant species are more important than other species for stabilizing community biomass. To our knowledge, the current study is the first to find that precipitation variability can destabilize rather than stabilize natural grassland communities by increasing species synchrony, suggesting it may be responsible for the impaired stabilities of communities and vegetation activities under high precipitation variability in previous region-scale investigations [6,47]. Our results contrast with the results of a recent meta-analysis of nine sites across grasslands of North America, which found that there precipitation variability stabilized communities by promoting compensatory dynamics [27]. This discrepancy may be due to drier average conditions of the 23 sites here studied across the Inner Mongolian grassland (precipitation ranged from 186.2 to 398.0 mm), leading to a positive correlation between precipitation and biomass for all or at least most species [41], whereas under wetter average conditions in North American grassland (precipitation ranged from ca 250 to ca 900 mm at the sites analysed in [27]) some species may actually increase in biomass in drier than average years, thus maintaining more constant community biomass. Therefore, the pattern observed in the present study may also be characteristic for even drier regions with semi-arid and desert ecosystems. In addition, the current study is also the first to quantify how different species-abundance groups affect species synchrony and found that here only dominant species impacted species synchrony significantly (electronic supplementary material, figure S8), potentially owing to differences in their responses to environmental fluctuations (electronic supplementary material, figure S3). Recent theoretical analyses have indicated that high unevenness of species abundance distributions can weaken the dependence of species synchrony on species richness and cause it to be strongly driven by a few abundant species [13,25]. Such a theoretical prediction has been supported by long-term (more than 20 years) single-site observational studies in grasslands of Inner Mongolia and the Qinghai-Tibet Plateau showing that compensatory dynamics between key functional groups maintain a stable community biomass [8,17]. Our results support such a theoretical prediction as well. Thus, the current study not only suggests that the strong dependence of species synchrony on few abundant species may be general in natural ecosystems characterized by high unevenness, but also provides a tool to examine and quantify this dependence. In the current study, we investigated stabilizing effects of biotic mechanisms on temporal variation in plant community biomass. We did not quantify the effects of anthropogenic disturbances, e.g. grazing and mowing, although they can be large [16]. It is, of course, conceivable that anthropogenic disturbances can reduce the richness of dominant species within grassland communities and affect other variables in the systems depicted in our SEMs, thereby indirectly affecting grassland stability in Inner Mongolia. Indeed, Inner Mongolian grassland has been seriously disturbed by livestock overgrazing and coal mining during past decades [37,45,61]. Furthermore, predictions for the future climate include increased precipitation variation [42,43], which could decrease grassland stability, as shown in the present study. The datasets and R code used for this study can be obtained from the Dryad Digital Repository: https://doi.org/10.5061/dryad.ht76hdrc5 [62]. Y.W., X.N., L.Z., C.L. and W.M. designed the study. Y.W., X.N., L.Z., C.L., B.M., Q.Z., J.Z. and W.M. compiled the data. Y.W., B.S. and W.M. produced the results and wrote the manuscript. We declare we have no competing interests. This study was supported by the National Nature Science Foundation of China (grant nos 31370454, 31960259, 31971434 and 31600385), the National Key Research and Development Program of China (grant no. 2016YFC0500602), the Ministry of Science and Technology of China (grant no. 2015BAC02B04) and the Natural Science Foundation of Inner Mongolia (grant nos. 2019MS03089, 2019MS03088 and 2015ZD05). B.S. was supported by the University Research Priority Program Global Change and Biodiversity of the University of Zurich. The authors thank Zhuwen Xu, Shaopeng Wang and Jens-Christian Svenning for their constructive comments. FootnotesElectronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.4969886. References
Page 5As we arrive at the 2020 expiration of the Aichi Biodiversity Targets under the Convention on Biological Diversity (CBD), the continuing loss of biodiversity remains a seemingly intractable environmental challenge [1] with grave implications for human wellbeing and the supply of valuable ecosystem services [2]. Some 322 vertebrates have become extinct since 1500, and more than 27% of all assessed extant species are threatened with extinction [2,3]. At a global scale, the average abundance of monitored vertebrate populations declined by 60% between 1970 and 2014 [4]. With the average rate of vertebrate species loss over the last century being up to 100 times the background rate, there is little doubt that we have entered an era representing the sixth mass extinction [1]. Deforestation has been a significant driver of this worldwide biodiversity crisis. Over a century ago, most clearance was of temperate forests [5], leading to observed species extinctions [6], while in the last decades, the main deforestation frontiers and risks to biodiversity have been in the tropics [7,8]. Tropical forests are some of the most biodiverse ecosystems on Earth, harbouring over half the world's terrestrial species [9]. Yet, deforestation of tropical forests, reducing their land coverage from 12% to less than 5% [10], along with their degradation and fragmentation, have resulted from large-scale industrial and local subsistence agriculture [11] as well as logging, fires and road building [12]. This represents a loss of important resources and habitat for humanity (between 1.2 and 1.5 billion people are directly dependent on ecosystem services provided by tropical forests [13]) as well as biodiversity, with far-reaching implications for the climate system [14] and global carbon cycle [15,16]. Land-use change is predicted to continue as a major driver of terrestrial biodiversity loss for the rest of this century [17]. In order to assess the impacts of land conversion pressures, it is crucial to develop national-to-global scale biodiversity measurements [18]. Owing to the importance of forests as habitat for many species, forest area is often employed as an indicator in global agreements and processes aimed at slowing and reversing the decline of biodiversity. Under the Strategic Plan for Biodiversity 2011–2020, for example, Aichi Target 5 focuses on halving the rate of loss of forests and other natural habitats by 2020 [19]. The suite of indicators for Sustainable Development Goal (SDG) 15 (Life on land) of the 2030 Agenda on Sustainable Development includes forest area as a proportion of total land area, and the proportion of forest and other ecosystems covered by protected areas [20]. Similarly, indicators used to monitor biodiversity conservation in the Forest Resources Assessment of the Food and Agricultural Organization (FAO) comprise area of primary forest, forest area designated for the conservation of biodiversity and forest area in legally established protected areas [21]. However, the pertinence of forest area as a relevant indicator of forest biodiversity has never been tested at a global scale. While habitat loss is the major driver of forest biodiversity loss, a focus on forest area alone risks masking other pressures on forest vertebrates that can operate below the canopy in conjunction with or independently of forest cover change. Consequently, areas with stable or increasing forest cover might be experiencing undetected declines in forest vertebrates, leading to the so-called empty forests that appear intact but have lost many of their large animals [22]. Understanding the status of forest biodiversity is important not only for species conservation but also because biodiversity loss can have consequences for forest health [12,23] and carbon stocks [24,25]. The status of the world's forests is a critical factor in the avoidance of dangerous climate change, with afforestation or reforestation being critical to many of the scenarios consistent with meeting the 1.5°C target [26]. Concurrently, the conservation of biodiversity in forests can have direct carbon benefits. Forest vertebrates, particularly large birds and primates, play an important role in forest regeneration and long-term carbon storage [27]. A loss or reduction in forest vertebrates from regions with a high proportion of large-seeded animal-dispersed tree species, such as Africa, Asia and the Neotropics, can lead to carbon losses in forests [24,25,28]. Defaunation therefore threatens the role that forests play as essential carbon stores and sinks, risking the investments made by governments and non-state actors in forests as carbon ‘banks’. Using the Living Planet Index (LPI) methodology [29,30], we aimed to develop the first global indicator of forest vertebrate specialist populations to improve assessments of forest biodiversity status. Given the decline in area of natural forest over time [31] and the link between habitat loss and biodiversity loss [32], we expected to find that forest vertebrates were in decline. We then assessed whether trends in forest vertebrate populations were related to changes in tree cover, derived from satellite-derived tree cover datasets that matched the forest vertebrate data in space and time. If tree cover were a good indicator of forest biodiversity, we would expect to find a positive relationship between forest vertebrate population change and tree cover change. We therefore tested two hypotheses:
The Living Planet Database (LPD) contains time-series abundance data for over 22 000 vertebrate populations including more than 4200 species across the globe, with the earliest records dating back to the 1950s (www.livingplanetindex.org). The data are collated from a range of sources, including peer-reviewed literature, grey literature, online databases and data holders. Metadata associated with each population, such as taxa, region, biome or habitat association, are also entered into the database. The decision to develop an indicator for forest specialists as opposed to all forest species follows the approach, but not the same method of selection, of the indicators developed for European birds [33]. Given that specialists depend entirely on forests, their use in this indicator would provide a better representation of ecosystem health. We defined forest specialists using the habitat coding from the IUCN Red List [3]. Those with ‘Forest’ listed as one of multiple major habitats for that species were considered forest generalists, while those with only ‘Forest’ listed as the major habitat were considered forest specialists. This definition of specialist is narrow as the ‘Forest’ category from the IUCN Red List refers to natural habitat and does not include artificial habitats such as plantations. However, as the category applies to the major habitats a species occurs in, it is still possible that all or part of a population may be located in or adjacent to a plantation. The forest specialists dataset comprised 268 forest specialist species (455 populations): 135 birds, 89 mammals, 19 reptiles and 25 amphibians. See electronic supplementary material, S1 for a breakdown by realm and taxonomic class. We followed the approach of the diversity-weighted LPI [30] to create a weighted index proportional to the species richness of each biogeographic realm and taxa in the dataset, and also to enable results to be compared with the global terrestrial LPI. In order to calculate weightings for each taxon and realm, the total number of vertebrate species from each taxonomic class and biogeographic realm that have ‘Forest’ listed as a habitat was taken from the IUCN Red List. Unlike for birds, mammals and amphibians, the coverage of reptile assessments in the IUCN Red List is not comprehensive so we did not have a full list of forest reptile species globally. However, the number of forest reptiles by realm was considered usable, given that the proportion of reptile species in each realm was similar to amphibians and also because spatial patterns of species richness tend to be similar among other vertebrate groups [34]. To create the subsets for the indicator, we disaggregated the data according to three taxonomic groups (mammals, birds, herptiles) by five realms (Nearctic, Palaearctic, Neotropical, Afrotropical, Indo-Pacific). Combining amphibians and reptiles into a herptile group, and Indo-Malaya, Australasia and Oceania into a single Indo-Pacific realm was a response to low data availability for these subsets. The final combinations yielded a total of 14 subsets as there were no time-series data available for Palaearctic herptiles. The Forest Specialist Index was calculated using the R package rlpi (https://github.com/Zoological-Society-of-London/rlpi) following the approach in McRae et al. [30]. The weightings calculated above for forest species were applied to each of the 14 subsets. In order to examine trends within these subsets of the data and by forest biome, we compared the mean and standard error of the species trends within each of the subsets. The individual species trends were available as one of the outputs of the rlpi package. While the Forest Specialist Index reflects population changes in forest specialists to more accurately reflect ecosystem health, changes in tree cover may also affect populations of forest generalists. We, therefore, selected all forest specialists and generalists that were surveyed at a specific location (defined as a discrete area such as a national park or sample area of a forest; a non-specific location comprises a larger survey area such as a province or country). For each population, the period encompassing the first and last year of survey data is subsequently referred to as the study period. Many population records do not have data available for every year of the study period. We determined annual predicted abundance values per population by fitting generalized additive models (GAMs) to the time-series population data where survey data were available for at least 6 years, and linear regressions where data were available for between 2 and 5 years, following Spooner et al. [35]. In order to assess the relationship between tree cover change and forest vertebrate populations, we required a continuous measure of tree cover spanning multiple years and at a resolution that is sensitive to the local changes that are likely to be relevant to populations. Various global datasets exist that provide continuous tree cover values for multiple years and vary in tree cover definition, spatial resolution, temporal coverage and frequency (electronic supplementary material, S2). Currently, the highest resolution global datasets (e.g. approx. 30 m) are available for a shorter temporal coverage than some datasets with a coarser resolution. Higher resolution datasets allow more fine-scale detection of changes in vegetation cover, while longer-term datasets increase the likelihood of detecting a relationship between tree cover change and population change by increasing the number of populations and years that can be analysed. We opted to run our analyses twice, once using the shorter-term (2000–2017) 30 m Landsat Global Forest Change dataset (hereafter referred to as the Hansen dataset; [36]) and once using the longer-term (1982–2016) 5.6 km MEASURES VCF5KYR dataset, which includes annual fractional tree cover and bare ground cover values (hereafter referred to as the Song dataset; [8]). In addition to fractional tree cover in 2000 and 2010 (2010 layer accessed from [37]), the Hansen dataset provides annual tree cover loss as a binary presence/absence value for 2000–2017, defined as complete stand replacement or a change from a forest to a non-forest state within a pixel. This information allows the estimation of deforestation rates, but may mask fine-scale changes within pixels such as a reduction (but not complete loss) in tree cover and assigns gradual losses that occur over multiple years to a single year. It is important to note that, while the 30 m dataset used in these analyses comes from the Global Forest Change dataset, neither this nor the Song dataset differentiate between natural, semi-natural or non-natural forests (such as plantations). Thus, while losses (or gains) in tree cover might reflect deforestation (or regeneration) in natural forests, in plantations, this might reflect harvest (or growth) of products grown specifically for human extraction that may provide lower quality habitat for forest vertebrate populations. Systematically collected global data on tree plantations are lacking. The Global Forest Watch (GFW) Tree Plantations layer records tree plantations in a single year (2013/2014) for only seven countries [38] and is, therefore, unsuitable for our analyses. A recently released near-global dataset on plantations by GFW [14] is also unsuitable, as the reference year is 2015. In the absence of suitable global information distinguishing natural and planted forests, we, therefore, refer to tree cover rather than forest cover whenever discussing values derived from the spatial tree cover datasets used in this analysis. We fitted a 5 km radius around each population, based on the mean range size across all forest populations (electronic supplementary material, S3), and extracted annual tree cover area and bare ground area for 1982–2016 using the Song dataset and tree cover area in 2000 and 2010 using the Hansen dataset. We additionally extracted annual loss values for 2001–2017 from the Hansen dataset, using per-pixel tree cover in 2000 to estimate how much tree cover was lost per buffer per year. All data extraction was carried out in Google Earth Engine [39]. We plotted annual tree cover values from the Song dataset against year to visually assess temporal changes in tree cover per location. We identified substantial inter-annual fluctuations in tree cover at some locations that were unlikely to reflect true changes. To smooth these fluctuations in the Song dataset, GAMs were fitted to the annual tree cover values within each buffer to obtain annual fitted tree cover values. We reduced the annual fitted population data to only include years that fell within 1982–2016 when analysing the effects of tree cover change with the Song dataset and 2000–2015 when analysing with the Hansen tree cover dataset. In both cases, we removed populations that no longer had greater than or equal to 2 years of data spread over at least a 5-year period (electronic supplementary material, S4 and S5). Using the annual logged values from the GAM and linear regression performed earlier, we calculated an average rate of change value per each remaining population as our response variable, following Spooner et al. [35]. Using the Song dataset, we reduced the annual fitted tree cover values to match the study period of each population, with a 1-year lag (i.e. tree cover in year t matched to population data in year t + 1). We then calculated three predictor variables from the fitted tree cover values: mean tree cover during the study period; mean bare ground cover during the study period; and the tree cover trend over the study period, taken as the year coefficient from an ordinary least-squares regression of annual fitted tree cover on year. We also calculated three predictor variables from the Hansen dataset: tree cover in 2000; the area of tree cover lost over the study period (based on loss data only); and the proportional change in tree cover between 2000 and 2010 (as these are the two years with percentage tree cover per pixel available). We removed populations with zero tree cover in all years from the analyses, leaving 1668 generalist and 175 specialist populations in the analyses using the Song dataset compared with 685 generalist and 74 specialist populations in the analyses with the Hansen dataset (see electronic supplementary material, S3 and S4 for a breakdown by realm and taxonomic class, respectively). Fewer populations were included in the analyses with the Hansen dataset because the shorter temporal period covered by the Hansen dataset (2000–2015) meant fewer populations had data overlapping that period, compared with the longer-term Song dataset (1982–2016). In order to examine the agreement between the two tree cover datasets, we calculated tree cover change per population from 2000 to 2010 using values derived from the Song dataset and from the Hansen dataset. We then assessed the correlation between the two sets of tree cover change values for the 685 populations included in the Hansen analyses. The correlation between the two datasets was highly significant but had a low correlation coefficient (Pearson correlation coefficient = 0.171; p < 0.001). This is in agreement with other studies that have found discrepancies between tree cover datasets when assessing tree cover change or area [40,41]. Forest vertebrates are affected by many drivers that may occur independently of, or in conjunction with, tree cover change. We selected correlates for our analyses through a literature review and information stored in the LPD, which includes any threats specified by the source of the population data. Exploitation, including the hunting, persecution, indirect killing or collection of wild individuals for trade, is likely to be a key driver of some forest vertebrate populations [42]. We, therefore, included in our analyses a binary variable specifying whether the primary threat to the population was or was not exploitation. It is possible that body size may impact species' sensitivity to forest change [43]. To investigate this effect, we took adult body mass values per species from the Amniote [44], AmphiBIO [45] and EltonTraits 1.0 [46] databases. Where species-level body mass information was not available, we assigned the species the mean body mass of its genus, family or order (higher taxonomic ranks used where data were unavailable for lower ranks). The body mass values were log-transformed (base 10) to normalize them. We calculated the density of roads within the study area, defined as the total length of roads within each population's 5 km buffer, using the gROADS v. 1 dataset [47]. We used the UN-Adjusted Gridded Population of the World V. 4 dataset [48] to calculate the mean human population density (HPD) within each buffer in the year 2000. Finally, we calculated the mean travel time to the nearest city or densely populated area for each buffer from the Accessibility to Cities 2015 dataset [49]. At some locations, multiple populations were monitored over the same period, so we chose to fit a model to the data that would take into account their non-independence. For each predictor variable, we fitted mixed effects models using the ‘lme4’ package [50] with the average rate of change of each population as the dependent variable, location as a random effect and the predictor as a fixed effect. We fitted separate models for each predictor variable to identify any relationships between these variables and population change, with the aim of fitting multivariate models where evidence of a relationship was found for more than one predictor variable. To determine whether a predictor variable was a significant driver of population change, we calculated Akaike's information criterion (AIC) for all models and compared them with the AIC of the null model including only a random effect of location. We considered a predictor variable to have significantly improved the model fit if inclusion of the variable lowered the AIC by at least 2 compared with the null model (a more negative AIC indicates a better model fit; [51]). We fitted these models to all forest populations (generalists and specialists) and additionally to forest specialist populations only. All analyses were carried out in the statistical software R v. 3.5.1 [52]. We investigated whether any groups of species were having a significant influence on the models. In the absence of any groups of influential species, models iteratively excluding one group at a time would not produce substantially different model estimates. We used the ‘influence.ME’ package [53] to produce estimates from models that iteratively excluded the influence of each genus, where each predictor variable was fitted in a univariate mixed effects model with genus as a random effect. We used the ‘sigtest’ function to test whether excluding any genus changed the statistical significance of any of the predictor variables in our models. We then examined the influential genus to determine the cause and, if the genus was known to be responding to a driver other than those included in our analyses (e.g. disease, poisoning), we repeated our analyses with the genus omitted. The Forest Specialist Index declined by 53% between 1970 and 2014 (figure 1a; index value: 0.47; range 0.30–0.73). This indicates an average decline in 455 monitored populations of forest specialists at an annual rate of 1.7% per year. By comparison, the terrestrial LPI declined by 41% between 1970 and 2014 (figure 1b; index value: 0.59; range 0.44–0.79), representing an average decline for 5175 monitored terrestrial populations with an annual rate of 1.2% per year. The decline in the Forest Specialist Index was steepest between 1970 and 1976. The percentage of all species that had an annual declining trend was consistently between 50 and 65% during the time period except for the late 1980s, early 2000s and 2013–2014, when the proportion dropped below half (electronic supplementary material, S6). These time periods are illustrated by corresponding changes in the index to a slower decline. There is an increase in the percentage of increasing annual trends in 2013 and 2014 and the percentage in 2014 is the highest out of all 44 years; this pattern is notable across all taxa (electronic supplementary material, S7). The average rate of change per species was negative for herptiles and mammals and slightly positive for birds (figure 2), with no overlap between the error bars of each group. This result was echoed when comparing declining and increasing years. There were more declining years than increasing among species trends for mammals (53% of all annual data points) and herptiles (63% of all annual data points); the reverse was true for birds, where there were more increasing years (52% of all annual data points). For all taxa, the percentage of increasing and declining annual trends varied across the time series (electronic supplementary material, S7). The average rate of change per species was negative for tropical realms and tropical biomes and positive for temperate realms and biomes (figure 2), with no overlap between the error bars for the two biome groups. Similarly, the number of declining species trends from tropical realms and tropical forest biomes was greater than increasing (electronic supplementary material, S8), while the reverse was true of temperate realms and temperate forest biomes (electronic supplementary material, S8). Figure 1. Weighted index of population change from 1970 to 2014 for (a) 268 forest specialist species and (b) 1853 terrestrial species (includes the forest specialist species). Solid line shows the weighted index values and shaded region shows the 95% confidence for the index. (Online version in colour.)
Figure 2. Total rate of change in forest specialist populations averaged by species, with standard error. Comparison by class (a), realm (b) and biome group (c). (Online version in colour.)
We identified one genus (Gyps) that had a large influence on the model estimates. Gyps vultures are a group of generalist species that have declined severely since the 1990s because of accidental poisoning from the veterinary drug diclofenac [54], and are, therefore, a very specific case that does not reflect responses of forest populations to any of the widespread pressures we have investigated. We, therefore, excluded Gyps vultures from our analyses. Mixed effects models including specialist and generalist forest populations and using the long-term Song tree cover dataset showed no evidence of a relationship between forest population change and tree cover trend (figure 3), mean tree cover, mean bare ground, exploitation, HPD, mean travel time or road density (electronic supplementary material, S9). Figure 3. Average rate of change of forest vertebrate populations (specialists and generalists) with abundance data covering at least a 5-year range between 1982 and 2016 from the LPD, and tree cover trend within a 5 km radius of each population's study location calculated over the same period as the population data from remotely sensed tree cover data [8]. Green, mammals; red, birds; blue, reptiles; black, amphibians. Circles, temperate biomes; triangles, tropical biomes. N=1668. (Online version in colour.)
We found a significant negative effect of exploitation on forest specialist population change, although this was based on exploitation being the primary threat to just 12 out of 175 forest specialist populations. We found no evidence of a relationship between forest specialist population change and any other predictor variable (electronic supplementary material, S10). Mixed effects models including forest specialists and generalists and using the Hansen tree cover dataset found no evidence of any relationships between population change and any predictor variables (electronic supplementary material, S11). We found no significant relationships between any predictors and population change when repeating the analyses using only forest specialist data (electronic supplementary material, S12). Our results indicate that the global abundance of forest specialists more than halved, on average, from 1970 to 2014. In context, populations of terrestrial species declined globally by an average of 41% over the same time period, which suggests that vertebrates in other terrestrial habitats have fared less badly. However, the population trends among forest specialists remain better than for species living in freshwater habitats, which exhibit more negative population trends [4,55] and a greater risk of extinction [56] than terrestrial counterparts. The result for the forest specialist index was consistent among mammals and herptiles but less so among birds, especially from temperate forests. Differences in average trends between taxonomic groups were significant and, while the effect of threats has not been quantified, the available evidence suggests the negative trend in mammals could be the result of targeted hunting, especially in the tropics [57]. The fungal disease chytridiomycosis, sometimes exacerbated by climate change, could explain the stronger negative result for herptiles (e.g. [58,59]). Abundance trends are worse in the tropics, as might be expected, given the more rapid rates of forest loss in tropical regions [21] over that period. The final years of the index, 2013 and 2014, showed an increasing trend as a result of a greater proportion of increasing annual trends among species than in previous years, across all taxa. As there have been other increasing trend years in the index throughout the time series followed by a decline (1991–1992, 2001–2002, 2004–2006), it is not possible to say at this stage whether the latest upturn in the Forest Specialist Index is a sign of a significant, longer-term increase in the abundance of forest specialists. In understanding the overall reduction in the rate of decline of the index after 2000, we need to consider three factors that are pertinent to interpreting trends in composite indices: species with increasing trends entering the dataset, species with declining trends leaving the dataset and improvement in species trends from declining to increasing or stable during this time period. The first two factors result from turnover in the species data that contributes to the index as data are not available for all 44 years for all species. This turnover in data is observed in our dataset: for example, between 2000 and 2002, data for 12 declining and four increasing species ended at the same time as data for 10 increasing and four declining species entered the dataset. This type of change in the dataset suggests that the reduced rate of decline may not entirely reflect overall improving status for species in the dataset, rather a change in the underlying data coupled with some species recoveries. This highlights a limitation of composite indices such as this where the temporal representation of species data is not comprehensive across the time series [60] and illustrates the need for diagnostics to accompany interpretation as well as additional data to strengthen the index. In addressing the third factor, and in order to eliminate any effect of data turnover, we looked at species with data present in all decades. These are predominantly bird species from the Nearctic, which are well monitored over the long term. After an initial decline, the average trend for this set of species does show an improvement to stability from the mid-2000s, but this trend is not yet increasing (electronic supplementary material, S13). The stabilization of trends in forest bird species in the Nearctic is consistent with other findings [61]. It is worth noting that species biodiversity data are currently skewed away from where species richness is greatest [62], limiting our ability to identify and address threats in some of the most biodiverse areas on the planet. The lack of population time series in the LPD from forest hotspots in Africa, Asia and the Amazon highlights this issue. To develop a more representative picture of the status of forest biodiversity and drivers of population change, these data gaps need to be filled. This will require greater investment in systematic, long-term, on-the-ground monitoring of forest vertebrates and improved data sharing within the research community. While remote sensing allows quantitative monitoring of forest cover change, limitations are to be expected in its use for monitoring forest populations: processes of defaunation are more cryptic and difficult to track [2], even occurring in large protected habitats [63]. The use of remote sensing to inform assessments of extinction risk for forest-dependent species has been demonstrated [64]. However, the relationship between habitat change and population change is not necessarily linear and the influence of threats other than habitat loss could also be important, which means that a species-specific approach may need to be taken when using habitat or land cover change to inform the status of a species [64,65]. Our results provide evidence that a satellite-derived assessment of forest cover change alone is inadequate as an indicator of trends in forest biodiversity. We did not find significant evidence of a consistent relationship between forest vertebrate populations and tree cover change in the surrounding area. Further, discrepancies between satellite-derived tree cover datasets in estimates of tree cover change or area indicate the uncertainties associated with tree cover assessments [40,41]. Analyses such as these would benefit from a global, systematically developed dataset categorizing forest areas into natural or planted forests, with temporal information detailing when each plantation was established. This would allow tree cover losses or gains within plantations to be identified, allowing for more rigorous checks of the relationship between populations of forest-dwelling species and natural forest cover change. Our finding of exploitation as a key driver of forest specialist population decline supports evidence presented elsewhere. An analysis of threat information for 8688 species on the IUCN Red List of Threatened Species identified overexploitation alongside agriculture (principally crop and livestock farming) as the main drivers of biodiversity loss [42]. The intensification of climate and other global environmental changes is predicted to interact with overexploitation and other pressures to lead to severe future degradation of tropical forests unless alternative, non-destructive development pathways are followed [12]. With most drivers of change interacting in space, time and organizational level [66], an explicitly linked set of forest biodiversity indicators may be more useful than reliance on any individual indicator to understand and communicate forest biodiversity trends and guide policy [67]. The Forest Specialist Index should be among such a set of indicators. This indicator has now been put forward through the Biodiversity Indicators Partnership to measure progress towards Aichi Targets 5, 7 and 12 (https://www.bipindicators.net/indicators/living-planet-index/living-planet-index-forest-specialists) and would complement existing indicators in monitoring progress towards SDG 15, the post-2020 framework under the CBD and in the delivery of the Paris Agreement. As such, it would also be a valuable inclusion in the Global Core Set of forest-related indicators as being coordinated by the FAO. The findings presented here also demonstrate the importance of complementing satellite-derived datasets with repeated on-the-ground species surveys and site-specific threat information when assessing the status and drivers of forest biodiversity, as advocated for elsewhere [68–70]. While remote sensing data have undoubtedly improved our ability to independently monitor and assess changes in forest cover, there are many additional drivers of forest population change that can only be identified by looking below the canopy. A focus on forest cover change alone risks masking below-canopy processes, such as defaunation, with grave consequences not only for forest biodiversity but also long-term forest health and carbon storage [24,27,28]. Therefore, we must not lose sight of the crucial role that site-level species monitoring plays in understanding trends and drivers of forest biodiversity change. The vertebrate population data were taken from the Living Planet Database which is hosted online at www.livingplanetindex.org. The data used for the analysis are available in the electronic supplementary material. Part of the dataset includes confidential data which have been shared under an agreement and are not publicly available. The species details, location and reference have been anonymized and the raw population data replaced with modelled population lambda values. The Forest Specialists Index was calculated using the R package rlpi available at https://github.com/Zoological-Society-of-London/rlpi. E.J.G. and L.M. carried out the statistical analyses with guidance from R.F., M.B.J.H. and S.L.L.H. W.B.-C. and W.D.S. conceived and coordinated the study. All authors contributed to the drafting of the manuscript. All authors gave final approval for publication and agree to be held accountable for the work performed herein. We declare we have no competing interests. The work of E.J.G., M.B.J.H., S.L.L.H. and W.D.S. was funded and supported by WWF-UK, WWF-Germany and WWF-France. Institutional support was provided by WWF-UK to the work of W.B.-C., and ZSL Institute of Zoology provided institutional support to the work of L.M. and R.F. We thank the following collaborators from WWF: Pablo Pacheco, Karen Mo, Lucy Young and Mark Wright for reviewing and discussing the development of this research, as well as Susanne Winter and Daniel Vallauri for supporting the research. We also thank Jack Plummer for conducting a preliminary analysis of the index whilst volunteering at ZSL, and Emma Martin at UNEP-WCMC for help with data collection. Footnotes†These authors contributed equally to this work. Electronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.4971446. © 2020 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original author and source are credited. References
Page 6The introduction of exotic species can cause substantial impacts on biodiversity and ecosystems [1–3], and such impacts are likely to exacerbate as rates of species introduction continue to increase [4,5]. Exotic predators are arguably the most disruptive group of introduced species [2,6], as they often exert impacts on native species far greater than those attributed to their native counterparts [7–9] and they are implicated in the extinction of hundreds of native species [1,2]. For instance, the accidental introduction of the brown tree snake (Boiga irregularis) onto the island of Guam, where there are no native arboreal vertebrate predators, caused the extinction of numerous species of birds, mammals, and reptiles [6]. The disproportionate impact of exotic predators on native communities is often attributed to prey naiveté—the failure of prey to recognize (or respond appropriately) to a novel predator species and/or the lack of an appropriate defence (sensu [10]). Such prey naiveté towards exotic predators likely derives from insufficient eco-evolutionary exposure [10–16]. For example, rats introduced to oceanic islands worldwide are implicated in numerous extinctions of mammals, birds, and reptiles that have no evolutionary experience with generalist mammalian nest predators [17]. However, rat impacts are reduced on islands that possess native rats or functionally similar land crabs, presumably because fauna on those islands are less naive to the effects of introduced omnivores [18]. Prey naiveté was originally conceived as a simplistic phenomenon where native animals become ‘easy prey’ to exotic predators owing to naive behaviour [11]. However, prey naiveté is now recognized as a more complex phenomenon and four levels of prey naiveté have been proposed [15,16,19]. Level-1 naive prey do not recognize the exotic predator as a threat, which precludes any antipredator behavioural responses [19]. Native animals experience level-2 naiveté if they recognize the exotic predator but show inappropriate antipredator behaviour [19]. Level-3 naive prey display an appropriate but ineffective behavioural response towards an exotic predator [19]. Lastly, level-4 naive prey over-respond to the exotic predator after experiencing excessive sublethal costs of predation [16]. In addition to exhibiting inadequate antipredator behaviour, prey species that lack evolutionary experience to exotic predation may also possess other morphological or physiological traits that make them susceptible to exotic predators such as insufficient armature, flightlessness, conspicuous scent, or inadequate camouflage [20]. Although prey naiveté is a well-accepted phenomenon [16], it varies under the influence of eco-evolutionary factors [14,15,21] whose relative importance and generality have yet to be quantified. We hypothesize that the occurrence and strength of prey naiveté stems from several, non-exclusive factors that can be clustered into four themes (table 1). First, prey naiveté can be promoted by persistent biogeographic (hence evolutionary) isolation between predator and prey [13]. The pronounced isolation of freshwater biota has been hypothesized to render prey more sensitive to introduced predators compared with terrestrial or marine biota [10,22] (Hypothesis 1 in table 1). Prey naiveté is also presumed to be more prevalent on islands than on mainlands [23–25], owing to lack of eco-evolutionary experience with exotic predators—or even native ones on predator-free islands (Hypothesis 2 in table 1). Likewise, predators introduced to geographically isolated or species-poor biotas are more likely to represent a novel archetype—that is, prey will not display antipredator responses towards exotic predators that are unfamiliar, where a practical proxy for ‘archetype’ distinction has been proposed at the taxonomic level of genus or family [10,16,26,27] (Hypothesis 3 in table 1). The introduction of a predator from a different biogeographic realm enhances the probability that the predator will be distinct from those of the recipient biota and thus unfamiliar [10] (Hypothesis 4 in table 1). The second theme is related to the way animals acquire antipredator responses (and lose prey naiveté) over time through adaptation, which could be a function of the number of prey generations since the introduction of a predator [28–30] (Hypothesis 5 in table 1). The third theme is related to the mediating role of latitude on prey naiveté, as novel predator recognition could be higher in low latitude communities, which generally experience greater and more diverse predation pressure [31–33] and thus whose prey may display antipredator behaviours to a broader variety of predator archetypes (Hypothesis 6 in table 1). Finally, the fourth theme is related to taxonomic specificity, as the recognition of introduced predators might vary across taxa [34], such that certain predators are more recognizable than others and certain prey are better adapted to recognize certain predators or entire suites of predatory taxa (Hypotheses 7 and 8, respectively, in table 1).
Many case studies suggest that these hypotheses are important predictors of prey naiveté [10,12,26,30,35,36], but no synthesis of global trends has been conducted. Here, we tested the generality of these eight hypothesized drivers of prey naiveté, with the goal of revealing which of these drivers can be used to effectively predict prey naiveté and conservation outcomes. We performed a search on 1 May 2019 following the guidelines of PRISMA (preferred reporting items for meta-analyses; [37]; electronic supplementary material, table S1). We entered the following terms in the Web of Science using the Advanced Search option: TS = (prey naiveté OR prey naivety OR naive prey OR lack of predator recognition OR antipredator behavio*) AND TS = (exotic OR invasive OR alien OR non-native), which produced 199 publications (electronic supplementary material, table S1). We also added 12 additional studies by examining the references of papers focused on prey naiveté (electronic supplementary material, table S1). Studies were included if they met the following criteria. First, each study empirically compared—in field or laboratory experiments—the behavioural response of prey to an exotic and a native predator. In this study, we only evaluated evidence of predator recognition (Level-1 prey naiveté [19]), which has been proposed as the most fundamental form of prey naiveté [38]. Second, the studies quantified behavioural responses and reported the mean, some form of variance (standard deviation, standard error, or confidence interval) and the sample size. Third, experiments within published articles were included as individual observations (i.e. number of rows on the database) if (1) investigators used different species of prey, native predator, and/or exotic predator, (2) experiments were performed with individuals from different locations, and (3) studies provided measurements for distinct behavioural responses to the same set of species of predators and prey because antipredator responses can be contrasting (e.g. prey might reduce activity in the presence of an exotic predator but not alter refuge use). Finally, to avoid temporal pseudoreplication, if a study measured a behavioural response through time (e.g. longitudinal studies), then the mean response over time was calculated. This criterion was adopted to better represent the generality of the behavioural responses. The effect size g was calculated as follows [39]: g=(XE−XC) JSDpooled, where XE and XC are the mean of the experimental (exotic predator) and control groups (native predator), respectively. J corrects for bias because of different sample sizes by differentially weighing studies as follows [40]:J=1–(3(4(NE+NC−2)−1)) . One can think of the effect size g as the difference in prey behaviour when in the presence of an exotic versus a native predator. Careful consideration was given when obtaining data from different metrics of predator avoidance, because the direction of the response variable depends on the specific behavioural response quantified. For instance, prey activity is a common metric of predator avoidance and decreases with increasing perception of risk, because prey are usually less active in the presence of a predator. On the other hand, refuge use—another common metric of antipredator behaviour—increases with increasing perception of risk. In order to standardize the direction of our metrics on antipredator behaviour, the data obtained from metrics that increased with increasing perception of risk were not changed and data obtained from metrics that decreased with increasing perception of risk were transformed to negative numbers (a negative symbol was added to the raw values for XE and XC). Therefore, g values near zero indicated predator recognition (e.g. prey respond similarly to an exotic and native predator), whereas values less than zero suggested prey naiveté (e.g. less perception of risk of the exotic predator than to the native predator), and positive values indicated prey perceiving an exotic predator to be more risky than a native predator. We obtained data from the text or tables of the studies or extracted measurements from figures in digital PDFs using ImageJ. The pooled standard deviation (SDpooled) was calculated as [40,41] SDpooled=(NE–1)(SDE)2+(NC–1)(SDC)2NE+NC−2 , where SD is the standard deviation of the experimental or control group and N is the sample size. When the standard error (SE) or the confidence interval (CI) was reported, the standard deviation was calculated. We weighted the effect sizes to account for inequality in study variance by using the inverse of the sampling variance, in which the variance for each effect size (Vg) was [40]Vg=(NE+NCNENC)+(g22(NE+NC)). There were four studies (see electronic supplementary material, table S2) that compared the behavioural response of an exotic prey to native and exotic predators and the native predators were considered to be the novel consumers of the exotic prey, therefore qualifying as tests of prey naiveté. Three papers that investigated antipredator responses to the exotic green crab Carcinus maenas in the North-Western Atlantic [42–44] were not included in the meta-analysis because the native prey Nucella lapillus is sympatric with C. maenas in the North-Eastern Atlantic (the native range of the green crab), and, hence, did not meet our criteria. In addition to the effect size, we recorded from each study the following factors: (1) ecosystem type (whether the exotic predator was introduced in a terrestrial, freshwater, or marine system), (2) insularity (whether the introduction was on an island or on a continental mainland including only data from terrestrial systems; Australia was considered a continental mainland, following [45]), (3) biogeographic realm difference (whether or not the location of introduction and the native range of the exotic species occupy the same biogeographic realm; terrestrial and freshwater systems were assigned to one of 11 biogeographic realms and marine ecosystems to 1 of 12—see electronic supplementary material, table S3), (4) taxonomic distinctiveness of the exotic predator (presence/absence of native predators in the introduced biogeographic region that belong to the same genus as the exotic predator), (5) the exotic predator taxonomic group (a posteriori categorized as six levels: fish, mammal, crustacean, herpetofauna, insect, and echinoderm), (6) the taxonomic group of the native prey (a posteriori categorized in six levels: fish, mammal, crustacean, herpetofauna, insect, and mollusc), (7) the number of prey generations since introduction (calculated by dividing the time passed since the exotic species was first recorded in the exotic region by the generation time of the prey species), and (8) the absolute latitude of the introduction of the novel predator measured in decimal degrees. We used the point of introduction instead of other potential spatial proxies—for instance, the midpoint of the full range of predator introduction—because the distributions of exotic predators and their native prey are usually patchy and often times unknown, and, more importantly, because many researchers used the patchiness of predator distributions to define the area of study (based on the presence or absence of the exotic predator). Meta-analyses were performed using the metafor package for R [46]. Treatments with less than or equal to 10 observations were dropped from the analyses to improve statistical robustness, which only included the removal of exotic echinoderms and insects (3 and 4 observations, respectively) from the analysis (see electronic supplementary material, table S4). We ran six independent mixed-effect models with different fixed predictors and in which ‘study ID’ and ‘experiment ID’ were always included as nested random factors, to account for multiple observations attained from the same study and experiment. In addition, the number of generations since introduction was added in each of the six models as a covariate to account for the potential effects of adaptation on prey naiveté through time. These six independent models included the following fixed, categorical predictors: (1) the system type of the introduction (with three levels: terrestrial, freshwater, or marine), (2) insularity (with two levels: mainland or island), (3) taxonomic distinctiveness of the introduced predator (with two levels: yes or no), (4) difference in biogeographic realm (with two levels: same or different if the biogeographic realm of the exotic predator was the same or different than the biogeographic realm of the introduction), (5) the taxonomic group of the exotic predator (with four levels: fish, mammal, crustacean, and herpetofauna; insect and echinoderm were excluded because they had ≤10 replicates), and (6) the taxonomic group of the native prey (with six levels: fish, mammal, crustacean, herpetofauna, insect, and mollusc; some important taxa (e.g. birds) were not included as no publications were found that met our criteria for comparing the response of these groups towards native and exotic predators). Effect sizes were considered significant if the 95% CIs did not overlap with zero. We also ran two further independent mixed-effects models with the nested random factors described above and a continuous fixed factor: the number of prey generations since the introduction of the novel predator, and the absolute latitude of introduction. For these models with continuous predictors, their significance was determined by the p-value of the moderator [46]. Publication bias can distort the results in a meta-analysis [40] by, for instance, overestimating prey naiveté towards exotic species. The functions regtest and trimfill are not implemented in the metafor package for mixed-effects models. Therefore, potential publication bias was evaluated using Egger's regression test [47] by running models that included the standard error of the effect sizes (included as the square root of the variance) as a moderator [48]; bias was determined when the intercept of the model was different from zero at p-values ≤ 0.05. In addition, we examined the data for potential outliers by looking at the effect sizes with standardized residual values exceeding the absolute value of three [49] using the rstandard function in R. Adjusting for publication bias did not change the outcome of the analyses (by comparing fitted random-effects models with and without the influence of the potential outliers; electronic supplementary material, table S4), indicating minimal influence of potential outliers. We found 40 studies that met our criteria to be included in the final dataset (electronic supplementary material, table S2), which comprised a total of 214 observations. The studies were published between 1993 and 2018 (electronic supplementary material, table S2) and included 47 unique study locations of introduction (electronic supplementary material, figure S1). Overall, we included reports assessing prey naiveté in 61 species of prey, with 38 species of exotic predators and 57 species of native predators (electronic supplementary material, table S2). The majority of species of introduced predators in our study were from freshwater systems (54.6%; 117 observations out of 214) when compared with terrestrial (33.6%; 72 observations) and marine systems (11.7%; 25 observations). The models that included ‘number of prey generations’ had a lower Akaike's Information Criterion (AIC) than those that excluded this variable, so the variable was kept in the models, regardless of its significance. Naiveté was found to be significantly pronounced in animals from marine and freshwater systems (mean Hedge's g ± 95%CI = −0.79 ± 0.38 and −0.32 ± 0.25, p < 0.001, and p = 0.013, respectively; figure 1a) but not significant in terrestrial systems (g = −0.35 ± 0.42, p = 0.107; figure 1a). Likewise, significant levels of prey naiveté were exhibited by prey on islands (g = −0.31 ± 0.18, p = 0.001; figure 1b), but not by animals in terrestrial continents (g = −0.02 ± 0.15, p = 0.789; figure 1b). Prey naiveté was significant only when the original biogeographic realm of the exotic predator differed from the realm in which it was introduced (g = −0.47 ± 0.20, p < 0.001; figure 1c), but not if the introduction occurred within the same biogeographic realm (g = −0.30 ± 0.42, p = 0.165; figure 1c). Similarly, the taxonomic distinctiveness of the exotic predator in the introduced realm also predicted prey naiveté, as native prey were significantly naive to distinct exotic genera (g = −0.47 ± 0.21; p < 0.001; figure 1d), but not towards introduced species with a sympatric species in the same genus (g = −0.35 ± 0.37, p = 0.087; figure 1d). Figure 1. Determinants of prey naiveté. Influence of (a) system type, (b) insularity on terrestrial systems, (c) distinctiveness of the biogeographic realm, (d) taxonomic distinctiveness of the exotic predator (i.e. a congeneric species of the exotic predator does not exist within the recipient community), (e) exotic predator taxa, and (f) native prey taxa, which was assessed by comparing the behavioural response of native prey towards native and novel predators. Points indicate the mean effect sizes bracketed by 95% CIs estimated using mixed-effects models. Effect sizes less than zero indicate less antipredator response to a novel predator than to a native predator, and the opposite for effect sizes higher than zero. Effect sizes are considered significant if their 95% CIs do not overlap with zero. Number of observations used to calculate the effect sizes are indicated in parentheses.
We found significant evidence that two taxa of exotic predators (fish and herpetofauna—e.g. amphibians and reptiles) were not recognized by the native prey (g = −0.57 ± 0.23, and −0.53 ± 0.39, p = less than 0.001, and p = 0.007, respectively; figure 1e), whereas exotic mammals and crustaceans were recognized similar to native predators (g = −0.25 ± 0.37, and 0.008 ± 0.35, p = 0.193, and 0.962, respectively; figure 1e). We found significant evidence supporting that two taxa of native prey, herpetofauna and fish (g = −0.37 ± 0.34 and −0.60 ± 0.36, p = 0.036, and less than 0.001, respectively; figure 1f) were prone to be naive towards exotic predators, whereas species from four taxa (insects, molluscs, crustaceans, and mammals) did not exhibit overall prey naiveté (g = −0.42 ± 0.92, −0.32 ± 0.49, −0.41 ± 0.51, and −0.44 ± 0.46, p = 0.370, 0.202, 0.108, and 0.06, respectively; figure 1f). The probability of individuals expressing prey naiveté significantly decreased with the number of prey generations since introduction ((Q-test of the moderator (QM) = 4.332, p = 0.037; figure 2a). Prey species recognized novel predators as threatening as native predators after existing with the novel predators for an average of 215 prey generations, which coincided with the predicted 95% CI of the effect size overlapping with zero (figure 2a). The latitude of the introduction did not influence predator avoidance behaviour of prey to novel predators (QM = 0.287, p = 0.592; figure 2b). Figure 2. Determinants of prey naiveté. Influence of two continuous predictors: (a) number of prey generations and (b) absolute latitude of the introduction on prey naiveté. Solid line indicates the mean predicted effect sizes bracketed by 95% CIs (dashed lines) estimated using mixed-effects models.
Our meta-analysis supports the generality of some, but not all of our hypotheses concerning invasions, and yields some novel insights (table 1). As postulated, we found that prey naiveté was pronounced in freshwater systems but not in terrestrial systems. In concordance with the aquatic-terrestrial dichotomy hypothesis, terrestrial animals were rarely naive towards exotic predators. This phenomenon in terrestrial systems has been attributed to the homogenizing effects of historical biotic interchanges across land masses, whereas the persistent isolation of freshwater systems might have rendered them less experienced to a broader suite of predatory archetypes [10]. Unexpectedly, marine environments appeared to be the most susceptible to introduced predators, contrary to the expectation that they are similar to terrestrial continents in terms of biotic connectivity [10]. Reports of prey naiveté in marine systems are rare and we gathered information from seven publications that compared antipredator responses with exotic and native marine predators. Four of these studies investigated fish naiveté to the exotic lionfish Pterois volitans in the Caribbean, reporting consistent naive fish behaviour. Two other studies [50,51] investigated the exotic marine green crab C. maenas, which found support for predator recognition by native crabs and gastropods. The other marine study [52] reported a pronounced degree of prey naiveté towards the exotic seastar Asterias amurensis by native scallops in Australia. Therefore, although exotic lionfish might have skewed our overall findings on marine prey naiveté, results from the exotic A. amurensis support this trend and suggest that exotic predation threats in marine systems might have been underestimated by conventional wisdom as opposed to actual data. Evidence from this study supports the hypothesis that terrestrial animals on islands are generally naive towards exotic predators, representing the first global quantification of prey naiveté on islands. When isolated from predators, prey on islands can experience a rapid loss of antipredator behaviour through relaxed selection [53–56]. Indeed, some prey species lack predators in the isolated Galapagos Islands, which are often described as being naive to predatory risk [57]. Similar examples exist on less remote islands, such as snake-free Balearic Islands in the Mediterranean, where wall-lizards show a lack of antipredator behaviours such as tail-waving or slow-motion movement when exposed to introduced snakes [58]. Prey naiveté is a primary explanation for the more devastating impacts of introduced predators on oceanic islands compared with continental terrestrial systems [10]. However, only three studies in our database addressed prey naiveté on islands and two of those were coastal islands (the exception was New Zealand). We hypothesize that the degree of prey naiveté on remote oceanic islands likely exceeds that reported in this meta-analysis. Australia was included as a continental mainland in our study owing to its large size; we performed an additional test by including Australia in the island category, which did not change our findings (g = −0.16. ± 0.15.; p = 0.039 and g = −0.04 ± 0.15; p = 0.853 for island and mainland, respectively), suggesting that our results robustly support the hypothesis that terrestrial species on islands display pronounced levels of prey naiveté. Prey species adapt to predators by accumulating eco-evolutionary experience [21] that familiarizes them with a particular species or archetype—a set of predatory species that have similar morphological and/or behavioural adaptations to obtain prey [10]. Recognition of a novel predation threat by a native prey species depends in part on the degree of similarity between the exotic predator and native predators present in the invaded community [14,16]. Hence, differences in predator archetypes between the area of origin and the area of introduction of an exotic species can profoundly influence the degree of prey naiveté. In the present study, we tested two proxies of distinctive predator archetype (allopatric origin and generic distinctiveness of the exotic predator [59]) and both were related to prey naiveté. The response towards exotic predators was limited when the introduced predator belonged to a novel genus in the invaded community or originated in a different biogeographic realm. Our results support the hypothesis that predator archetypes might be limited to congeneric species, as suggested previously [27,35], but phylogenetic analyses assessing evolutionary distance between predator species would be warranted to test this hypothesis. These results also substantiate a statistical synthesis [59] showing that high-impact invaders, including predators, are likely to belong to genera not present in the invaded community, which expectedly occurs more frequently if the predator is native to a foreign biogeographic realm. Native prey appeared more likely to be naive towards reptile and fish predators. Indeed, many species from these groups have been implicated in extirpations and extinctions [60], although most attention has been given to iconic cases such as the Nile perch Lates niloticus [61] and the brown tree snake B. irregularis [6]. Native amphibians appeared to be sensitive to the introduction of predatory herpetofauna—mainly freshwater turtles and frogs (92% of the 24 observations)—where the majority of prey species were frogs (83% of 24 observations). On the other hand, exotic predatory fishes were represented broadly (17 freshwater and one marine exotic fish species with 78% and 22% of the 78 total observations, respectively) and their prey belong to four taxonomic groups (insects, fishes, herpetofauna, crustaceans), suggesting that the identification of exotic fish as a predation threat might be generally elusive. We performed an additional analysis to ascertain whether the high probability of herpetofauna and fish to encounter naive prey was due to taxonomic affiliation and not simply driven by ecosystem type (freshwater, terrestrial, marine). We found similar results for these two taxonomic groups, regardless of the ecosystem type (g = −0.39 ± 0.27; p = 0.005 and g = −1.07 ± 0.44; p < 0.001 for freshwater and marine fishes, respectively, and g = −0.39 ± 0.41; p = 0.060 and g = −1.29 ± 1.11; p = 0.022 for freshwater and terrestrial herpetofauna, respectively), supporting our findings of likelihood of prey naiveté towards fish and herpetofauna. A surprising result was that prey recognize exotic carnivorous mammals as a predation threat, despite that their exacerbated impacts have been commonly attributed to prey naiveté [23]. Similarly, a recent meta-analysis investigating prey naiveté towards exotic mammals in Australia found high-risk aversion towards canids: the European red fox Vulpes vulpes and the dingo/dog Canis lupus dingo/familiaris [38]. The majority (62%) of observations in our dataset involving exotic mammalian predators were for canids. When we re-ran our analysis excluding canids, prey were marginally naive to carnivorous mammals (g = −0.42 ± 0.49; p = 0.09). Thus, we speculate that the canid family, which has long been present in most continents (including Australia, where Canis lupus dingo was introduced 4000 years ago [62]), represents a predator archetype that could be more broadly recognized than many other archetypes, perhaps because of extensive evolutionary exposure associated with human domestication. Our findings suggest that fish, amphibians, and reptiles are generally more naive to exotic predators than mammals and invertebrates (crustaceans, insects, and molluscs) and thus likely more sensitive to the introduction of predators. Exotic species have been identified as the number one threat associated with the extinctions of herpetofauna worldwide according to the International Union for Conservation of Nature (IUCN) Red List [63], but the extent to which prey naiveté drove these extinctions remains to be determined in many cases. We did not find significant levels of naiveté for mammalian prey in general, although exotic species are also the most frequent threat recorded for their extinctions [63]. Similar to our findings, a recent meta-analysis indicates that mammals in Australia identify exotic foxes and cats as a predation threat [38]. The authors argue that despite this lack of prey naiveté (level 1 sensu [19]) the rampant decline of prey by exotic mammals in Australia [64] might still be driven by inappropriate or ineffective prey responses (levels 2 and 3 naiveté sensu [19]), which are rarely quantified. Remarkably, although prey naiveté is invoked as responsible for the strong ecological impacts of exotic species on birds [2,6,65,66], we did not find any papers that met our criteria for quantifying prey naiveté in birds, mainly because the few studies addressing prey naiveté in birds lack a comparative treatment with a native predator. Finally, fish do not appear to respond to the risk of predation by novel fish. Collectively, these findings suggest global patterns that could strengthen predictions concerning evolutionary exposure. The antipredator response of native prey to novel predators can evolve through time, if predation selects for predator recognition and avoidance behaviour [28]. Behaviours that determine the survival of individuals facing a novel predation threat can be subject to strong selection in the persistent presence of a predator [67]. If extinction is averted, evolutionary adaptation can be achieved in a small number of generations. For instance, the fence lizard Sceloporus undulatus acquired the capacity to avoid exotic predatory red fire ants Solenopsis invicta in North America within 40 generations [36]. Our meta-analysis shows that naiveté erodes with the number of prey generations following predator introduction, indicating a generalized pattern of adaptation [30,68]. Averaged across the various taxa in our study, approximately 200 generations are required for native prey to acquire an antipredator response towards exotic predators in the same manner as native predators. This phenomenon could explain, in part, the observed declines in negative ecological impacts of exotic predators over time. For example, the ecological impacts of the brown trout Salmo trutta—an exotic predator intentionally introduced globally—decreased linearly with time since introduction [69]. This reduction in the negative ecological impacts occurs circa one century after the introduction of the brown trout and it was hypothesized to result from either rapid evolutionary adaptation or prompt local extinction of native prey [69]. The capacity of prey to recognize exotic predators is conditional on the native prey averting extinction that often occurs before prey naiveté is assessed [1,2]. Our study might have underestimated the generation time required for predator recognition by omitting prey species that can never adapt or learn how to recognize exotic predators. We are aware of at least one extreme case in our dataset: several fishes exhibit limited antipredator behaviour in the presence of the exotic lionfish P. volitans in the Caribbean [35,70], where strong reductions [71,72] and even local extirpations [73] of fish populations have been reported. We also predicted that prey at lower latitudes would be less naive towards novel predators owing to the large suite of predatory species and relative high intensity of predation in the tropics [32,33]. Although our results suggest that novel predator recognition is not influenced by latitude, data from the tropics were limited—perhaps reflecting actual low numbers of successful introduced predators [74] or historical low sampling effort of non-native species in the tropics [75]. There are several potential limitations to the data included in this meta-analysis. First, studies were excluded unless they met several criteria, with the disadvantage of not considering the totality of evidence generated globally on prey naiveté. We only considered experimental designs that included empirical comparisons between native and exotic predators, to ensure a direct and consistent way to quantify the perceived risk threat of an exotic predator. Consequently, we excluded studies with controls such as ‘absence of exotic predator’, as those comparisons often require cautious interpretation (e.g. does the behavioural response of prey towards the exotic predator as compared with an empty control indicate predator recognition or simply a response to the presence of an organism, regardless if it is perceived as a predatory threat?). Second, our study only included measurements of level-1 prey naiveté (sensu [16,19]), which interprets a lack of response to an exotic predator (when compared with a native predator) as a lack of recognition of the exotic predator as a threat. However, native animals experience additional levels of naiveté (level-2, -3, and -4), which relate to appropriate, effective, and/or commensurate responses to exotic predators, respectively. Therefore, wildlife might still experience heavy predation by exotic predators despite low level-1 naiveté. By focusing on level-1 naiveté, our study did not consider physiological responses to the presence of predators [76], which can be considered another important form of prey naiveté. Finally, our dataset did not include a random subset of all exotic predators, which might have biased our results towards the most notorious (and presumably detrimental) of exotic species. Our meta-analysis identifies some global drivers of prey naiveté, paving the way for testing these drivers in different contexts. Assuming that prey naiveté results in increased mortality [14], our results point to several animal groups as being disproportionately sensitive to introduced predators. Some of these vulnerable cases were expected, such as insular terrestrial and freshwater fauna, whereas other cases were unpredicted, such as the high susceptibility of native prey to exotic predators in marine systems, or the vulnerability of specific prey taxa, including fishes and amphibians. The relationship between overall prey naiveté and the number of prey generations suggests that long-lived species could be particularly vulnerable to introduced predators. It remains to be determined how other eco-evolutionary factors influence the loss of prey naiveté through time—e.g. how does this rate differ across taxonomic groups and ecosystem types? Additionally, the most damaging groups of exotic predators were found to be animals that originate from a foreign biogeographic realm or that represent a new generic archetype. Particular attention should be given to the introduction of predatory fishes, reptiles, and mammals (perhaps with the exception of canids). This information could guide efforts to prioritize invasion threats to biodiversity and inform risk assessments of conservation schemes involving assisted colonization. Finally, we identified several areas in which the quantification of prey naiveté is notably scant (e.g. marine ecosystems, remote oceanic islands, and many common prey taxa) and these should be prioritized to clarify predictive patterns of prey naiveté. Our results support the view that prey naiveté is shaped by multiple eco-evolutionary factors [16,19,21,38]. The phenomenon is of increasing relevance to conservation, given that species introductions to novel ecosystems are accelerating globally [4], along with other forms of global change that might promote ‘disturbed predator-prey interactions’ (sensu [16]). For example, the poleward migration of species driven by changing isotherms [77], including the imminent arrival of unique shell-breaking predators in Antarctica [78], will add novel predator-prey interactions even into historically isolated regions. Therefore, we recommend that factors influencing prey naiveté be given explicit consideration in biodiversity risk assessments. Data available from the Dryad Digital Repository: https://doi.org/10.5061/dryad.gqnk98sjh [79]. Code availability statement: the R script used in this manuscript is deposited here https://github.com/antongamazo/Determinants-of-prey-naivete. A.A. and N.R.G. conceived the study, A.A., N.R.G., A.R., and J.T.A.D. designed the study, A.A. collected data, A.A. and N.R.G. performed the analyses, A.A. wrote the first draft of the manuscript with substantial input from all authors, A.A., A.R., and N.R.G. contributed extensively to revisions, and all authors approved it for publication. We declare we have no competing interests. The Natural Environment Research Council (NERC) and the Natural Sciences and Engineering Research Council of Canada (NSERC) supported this study with grants to J.T.A.D. and A.R., respectively. We thank the School of Biological Sciences at Queen's University Belfast, the Queen's University Marine Laboratory, Julia Sigwart, Christine Maggs, and Bernie Curran for logistical support and Daniel Barrios-O'Neill for providing valuable feedback on an early draft of the manuscript. FootnotesElectronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.4971209. © 2020 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original author and source are credited. References
Page 7The persistence of spatially distributed species depends on aspects of local population dynamics and on dispersal [1]. Spatial management of a species therefore needs to consider both processes simultaneously. The metapopulation paradigm, where local populations are viewed as relatively discrete spatial entities that interact through migration, has proven very useful in understanding the interplay between dynamics and connectivity (dispersal) in a wide range of species including plants, amphibians, insects, birds, fish and mammals [2–4]. As such, the metapopulation approach is being increasingly applied to management in both the terrestrial and marine environments [3,5]. The dynamics of metapopulations are influenced by natural and anthropogenic factors, such as density-dependent natal dispersal and migration rates; [6–9]; habitat connectivity, loss and fragmentation [9,10]; and environmental heterogeneity [7,11]. At a regional level, it is the balance between local births and deaths, combined with net migration, which drives local population dynamics and persistence. Highly variable habitat quality among patches can also lead to source–sink dynamics [6,12,13]. The key idea is that in good quality regions, mortality is lower than reproduction. Surplus individuals from these ‘source' populations emigrate to lower quality regions, such that even if mortality is higher than natality, these ‘sink' populations can persist. Source–sink metapopulations are of particular interest because they are very susceptible to the effects of localized but abrupt perturbations affecting source populations, which can lead to overall metapopulation decline and eventual extinction [14]. An important implication of metapopulation theory is that, in the absence of exogenous perturbations, a species may persist regionally despite some local population decay and extinctions. A balance between these local decays/extinctions and local growth/new colonizations is expected to maintain the overall metapopulation. Similarly, source–sink dynamics can support sink populations larger than their source over evolutionary timeframes [13]. Therefore, a local population decline or extinction may be simply the manifestation of normal metapopulation dynamics but may also indicate more widespread issues with metapopulation health, particularly when sudden local population declines involve previously stable or growing source populations. A prime example of changing dynamics in a metapopulation is the UK harbour seal (Phoca vitulina), which has been monitored for decades to provide regional population trends, local movement and genetic datasets [15–18]. The UK-wide abundance of harbour seals is currently 42 100 seals (95% CI: 34 500–52 300), which is comparable to the estimate 20 years ago at 45 550 (95% CI: 37 250–60 750) [17]. In contrast with this stable overall picture, there have been dramatic declines in abundance in key small areas (e.g. 95% decline: 2002–2017 in East Scotland) as well as in large regions, such as Shetland (40% decline: 2001–2006) and the North Coast & Orkney (85% decline: 1997–2016; figure 1). The reasons for these declines, while the populations around the majority of the UK are stable or increasing, are unknown. Factors currently being studied include increased indirect and direct competition (including predation) by grey seals [19] or other marine mammals and exposure to toxins from harmful algae [20]. Harbour seal populations exhibit a combination of structure and connectivity that make them suitable for metapopulation analyses. The species' central place foraging tactics mean that individuals generally feed within 100 km of sites at which they haul out (where they are counted) between foraging trips [16], yet there is evidence for large-scale movements between haul out sites, and between haul out and breeding sites, over 50 km apart [16,18,21,22]. Figure 1. Map showing the most recent harbour seal count (10 km2 resolution; [17]). The Seal Management Units considered in this paper (plus Northeast England SMU, with small population (max count less than 100) and for which we do not have any data) are shown as well as associated trends (line and associated 95% confidence intervals) in August counts (points; y-axis) as a function of year (x-axis), extracted from Thompson et al. [17]. (Online version in colour.)
Here we test the hypothesis that the observed regional declines in the UK harbour seal are part of the normal extinction–colonization dynamics of a single metapopulation, or alternatively, a response to a major perturbation driving changes in metapopulation dynamics. In doing so, we identify which local populations are likely to be sources or sinks. Ideally, metapopulation connectivity and source–sink dynamics would be determined using direct measures of demographic parameters and connectivity among and between local populations, respectively (i.e. survival, reproduction, recruitment and dispersal). However, estimates of these demographic parameters are not available for UK harbour seals and typically require datasets that follow individuals throughout their lifespan [23], precluding their estimation in a timeframe relevant for management of the current decline. In this context, there are two main difficulties that make assessing source–sink metapopulation dynamics in long-lived species challenging. First, there is no single approach that can determine whether or not movement of individuals among local populations contributes to local dynamics on the short timescales relevant for conservation. Satellite tracking data can be used to elucidate the level of movement between local populations but cannot determine if dispersing individuals leave descendants in the new location. Genetic data, on the other hand, can estimate per-generation migration rates representing real contributions towards local demography but for long-lived species these estimates may cover a period of several years and therefore may be too coarse-grained to detect sudden changes in migration patterns following a perturbation. Similarly, distinguishing between source and sink populations based on long-term population trends is unfeasible because a sink population may exhibit stable or even increasing census sizes due to the influx of migrants from a source population. On the other hand, genetic data can provide estimates of ‘retention' (proportion of individuals that remain in their local population), which to some extent are indicative of local recruitment but cannot determine if local birth rate exceeds local death rates, as expected in an ‘absolute' sink (cf. [24]). In order to make progress despite these challenges, we adopt a framework, bringing together genetic, location tracking and population trend data, to assess metapopulation identity and connectivity in addition to establishing source–sink dynamics. We first use population genetics approaches (genetic differentiation index) to establish if all local populations of harbour seals are members of the same metapopulation. Having established that South Eastern UK local populations belong to a metapopulation that extends beyond the British Isles we focus on those found in the North Western and North Eastern UK, all belonging to a separate metapopulation. Specifically, we characterize metapopulation connectivity and source–sink dynamics of local populations in two different time periods, before and after the start of the regional declines (henceforth referred to as pre- and peri-perturbation). Thus, we assess the degree to which local populations are demographically connected by estimating per-generation migration rates pre- and peri-perturbation using multilocus-genotype methods. We seek further support for these results using satellite tracking data providing estimates of short-term movement of adults (greater than 1-year old, non-pups) and pups. Next, we identify putative source populations based on genetic data as those that have both high internal recruitment and display emigration. We then use population trends to further support their ‘source' status as they should also be stable or growing. Finally, we use a similar procedure to identify putative sink populations as those that are net recipients of immigrants and still declining, in which case mortality is likely exceeding reproduction. The integration of the three data types––genetics, tracking and population trend––allow us to evaluate the viability of putative source and sink populations in the context of changes in connectivity pre- and peri-perturbation and overall trends in abundance. Specifically, we discuss whether or not local population declines and changes in migration patterns are consistent with overall metapopulation persistence or they are indicative of metapopulation decay and potential regional extinction. Our final dataset comprises microsatellite genotypes and animal location tracking data from 269 and 380 harbour seals, respectively (table 1). These data were collected from geographical units known as Seal Management Units (SMUs); 11 SMUs covering the UK were established using harbour seal haul out clusters identified from aerial surveys, and tracking and photo-ID studies [16,21,22]. Here, we primarily consider the SMUs that hold significant harbour seal populations (greater than 100 individuals counted on surveys; figure 1 and table 1). We discuss the results in the context of three different types of groupings: metapopulations, metapopulation subunits called local populations and SMUs.
The pattern of genetic diversity and differentiation, as well as tracking data, suggests that the UK harbour seal SMUs fall into two distinct metapopulations: a northern and a southern. First, the Southeast England SMU showed high and significant levels of genetic differentiation against all other UK harbour seal SMUs (FST > 0.2; electronic supplementary material (ESM), table S1). By contrast, Southeast England showed only weak differentiation against the European samples. The BayesAss results confirmed this, with estimates of recent migration between components of the two metapopulations typically being ≤1% (ESM, tables S2 and S3), consistent with demographic independence [25]. Therefore, we consider Southeast England and continental Europe part of one southern metapopulation, and all other SMUs part of a northern metapopulation (Northern Ireland and Scottish SMUs) and focus on the latter. There was significant genetic differentiation between most of the UK harbour seal SMUs within the northern metapopulation based on pairwise FST values, although this was not as substantial (FST: 0.02–0.14; ESM, table S1) as between the two putative metapopulations (FST: 0.18–0.30; ESM, table S1). The exceptions to this general pattern were that there was no significant difference between (i) West Scotland and the Western Isles SMUs, which were pooled to form a Northwest local population, and (ii) North Coast & Orkney SMU, and the neighbouring Moray Firth SMU, which were pooled to form a Moray Firth, North Coast & Orkney (MFNCO) local population. Thus, the FST estimates suggest a total of five local populations within the northern metapopulation: Northern Ireland, Northwest, MFNCO, Shetland and East Scotland. Different haul out sites within SMUs and across SMU subunits (e.g. south and central West Scotland) did not show significant differentiation (ESM, table S1). Although no genetic samples were available from Southwest Scotland SMU, we assume this SMU is part of the Northwest local population: there are similar population trends and no spatial differentiation in haul out clusters between Southwest and West Scotland SMUs (figure 1). The discriminant analysis of principal components (DAPC) clearly separated the northern metapopulation SMUs from Southeast England and continental Europe along linear discriminant 1 (figure 2). The SMUs within the northern metapopulation were characterized by a pattern of isolation-by-distance along linear discriminant 2, but also indicated some east–west division. The isolation by distance was confirmed by a statistically significant (p < 0.001) correlation between FST and ‘at-sea' distances between haul sout sites (ESM, figure S1). Figure 2. Individual genotypes plotted by LD from the discriminant analysis of principle components conducted with samples grouped by SMU and haul out site (latter in brackets). Mean values for each sampling partition shown by triangle. (Online version in colour.)
The distinctiveness of the northern and southern metapopulations was also supported by tracking data from 334 non-pup seals, showing no movement between metapopulations, but some movement within (ESM, tables S4 and S5). For instance, of the 83 individuals tagged in Southeast England, two travelled to the continent but none to the northern metapopulation. Likewise, in the northern metapopulation, movements were detected among several haul-outs and SMUs, as described below, but there was no movement to the southern metapopulation. Overall, these results provide strong support for defining a northern metapopulation, which excludes the Southeast England SMU. We assessed connectivity of local populations and SMUs from genetic, movement and demographic perspectives, using the microsatellites, location tracking and population trends, respectively. As the Southeast England SMU and European samples appear part of a distinct metapopulation, we consider here a BayesAss analysis with the harbour seal dataset covering only the northern metapopulation (ESM, table S6). Overall, migration connects the northern local populations, however, in many cases connectivity has sharply reduced in the peri-perturbation generation compared with the pre-perturbation generation. This is evidenced by a lack of first-generation migrants, but a substantial proportion of individuals with migrant ancestry, during the peri-perturbation period (i.e. descended from parents that migrated before the perturbation). Median estimates of recent gene flow (past two generations), inferred from the BayesAss analysis, are shown in ESM, table S6. Convergence was shown by traces and similar results across independent runs, with effective sample sizes for each parameter greater than 250. The Northern Ireland and Northwest local populations are highly connected, as shown by the high proportion of immigrants from the latter (0.26, 95% HPD:0.19–0.31) and the substantial degree of recent migration based on individual ancestry. By contrast, the Northwest local population appears to be mostly local recruits, based on individual ancestry data and the high proportion of non-migrants (0.91, 95% HPD:0.85–0.97). There is indication that the MFNCO local population has been a source of migrants to the Northwest local population based on ancestry and migration rates (0.05, 95% HPD:0.01–0.12), but with a decline from four likely second-generation migrants to two first-generation migrants in the past two generations. This suggests a decline in migration peri-perturbation compared with pre-perturbation. The MFNCO local population also has a high proportion of non-migrants (0.95, 95% HPD: 0.88–0.99) and no evidence of immigrants in the past one generation. By contrast, there is evidence of past gene flow from the East Scotland local population based on the migrant ancestry of two individuals. Furthermore, the MFNCO local population is contributing migrants to both the Shetland (0.20, 95% HPD: 0.11–0.29) and East Scotland local populations (0.09, 95% HPD: 0.01–0.18) based on migration rates over the past two generations. For East Scotland, the gene flow from the MFNCO local population in the current generation shows a decrease from the previous generation, again supporting a decline in migration peri-perturbation compared with pre-perturbation. For Shetland, there is also evidence of migration from the Northwest local population, as indicated by individual ancestry data and relatively high migration in recent years (0.08, 95% HPD: 0.01–0.16), despite a moderate and significant FST value of 0.07 between the two regions. The location tracking data of seals (non-pups) in the non-breeding season broadly supports the distinctiveness of the five local populations; in total, 21 (6.3%) tagged seals moved between SMUs, of which only one (less than 1%) moved local populations. None of the individuals tagged in Northern Ireland (n = 33), Shetland (n = 14) or East Scotland (pre or peri-perturbation; n = 33) moved between local populations. There was movement between the three SMUs comprising the Northwest local population, particularly between West Scotland and Southwest Scotland (ESM, table S5). There was also significant movement within the MFNCO local population, from the Moray Firth to North Coast & Orkney (n = 4/39), with 1/53 (1/34 tagged in southern Orkney) going in the opposite direction. The only movement between local populations was of one individual tagged in 2003 in northern Orkney, MFNCO local population (of 19), which moved to Shetland. Pup-tracking data indicated a higher degree of connectivity between SMUs and local populations compared with non-pups, as expected from the dispersing demographic class; in total 11/46 (26%) moved SMUs, of which nine (19%) changed local population. Of the 24 pups tagged in Lismore, West Scotland SMU in the Northwest local population, four (17%) moved SMUs within the local population (two to Southwest Scotland and two to Western Isles) and one (4%) moved to Ireland. The latter pup moved to a region from which we do not have any genetic samples so cannot assess the location's position in the metapopulation. However, as the seal moved to within 50 miles of the Northern Irish border, its movement could represent a shift between local populations. Of the 22 pups tagged in Orkney, MFNCO local population, seven (27%) moved to Shetland and one (4%) moved to West Scotland (Northwest local population). We looked at demographic connectivity by assessing published trends in population trajectories and identifying which SMUs have similar trends. The Northwest local population SMUs have shown stable population trends over the time period covered by the study (Western Isles, West and Southwest Scotland). However, Northern Ireland, a separate local population, has exhibited a constant gradual decline (figure 1 and table 2). The SMUs in the MFNCO local population underwent a sudden change in dynamics in the early 2000s. For instance, North Coast & Orkney were stable until 2001, whereas the subsequent survey in 2006 showed a dramatic decline in abundance. Moray Firth stabilized in the early 2000s after a period of decline. Both East Scotland and Shetland underwent a similar change in dynamics in the early 2000s, but where Shetland looks to be stabilizing at a depleted level, East Scotland continues to decline.
The combined analyses allow us to assess the levels of movement, as well as genetic and demographic connectivity, within the metapopulation on different timescales and using different aspects of the harbour seal's biology (figure 3 and table 2). On the UK west coast, our data suggest that the Northwest local population (Southwest Scotland, West Scotland, Western Isles SMUs) is a source population, as it has a high level of retention and substantial emigration to Northern Ireland (table 2). Furthermore, population abundance for Northern Ireland has been steadily decreasing over the study period, despite receiving immigrants from the Northwest local population, suggesting that it is a sink population. Figure 3. Inferred source–sink dynamics of UK harbour seal northern metapopulation. Black lines delineate the SMUs and coloured lines indicate inferred local populations, with arrows indicating movement from putative sources to putative sinks. Dots represent approximate locations of telemetry tag deployment and/or genetic sampling. (Online version in colour.)
We also hypothesize that the MFNCO is a source population. This local population appears to have a high level of internal recruitment, based on the BayesAss analysis. Furthermore, the genetic and pup movement data suggests that there is emigration from MFNCO to other local populations. In particular, it seems that MFNCO is a likely source population to both Shetland and East Scotland. Both latter two local populations have a substantial proportion of individuals with MFNCO ancestry based on BayesAss, and the pup-tracking data showed considerable movement from the MFNCO to Shetland (table 2, ESM, table S5). Finally, MFNCO, Shetland and East Scotland share similar population trends, in that they have dramatic declines subsequent to the perturbation. In previous generations, there appears to have been migration from East Scotland into both Shetland and the MFNCO local populations; however, there is little evidence of this in the past one generation spanning the perturbation. This decline in emigration coincides with the precipitous decline in the East Scotland local population. Overall, the results show that the northern harbour seal metapopulation is highly connected and contains two probable source local populations: MFNCO and Northwest. The decline in the MFNCO local population appears to have reduced connectivity between these local populations as well. For example, there are four individuals (4%) from the Northwest local population (n = 90) that are most likely second-generation migrants from the MFNCO local population, but only two individuals are likely to be first-generation migrants. This could represent a decline in connectivity over the past one generation during which the MFNCO population exhibited a pronounced decline. A key tenet of the metapopulation paradigm is that local population decay and decline do not necessarily threaten metapopulation persistence if there is a concomitant balance with new colonization and growth. Such balanced patterns of extinction and colonization have been empirically shown in butterflies Melitaea cinxia [26] and Speyeria nokomis apacheana [27], as well as the American pika Ochotona princeps [28]. Here we have shown that the northern UK harbour seal metapopulation has been subject to a recent perturbation that has impacted local population connectivity in a way that appears to go beyond the expectations of a simple extinction–colonization equilibrium. The disruption of migration we see at a local level seems to have wider impacts on the metapopulation, rarely before seen in empirical studies. This change in connectivity could eventually lead to genetic isolation and genetic erosion over time [29], which can be very difficult to detect in the short term given the long generation time of the species [30]. Predicting long-term consequences, therefore, requires the use of a thorough population viability analysis integrating both local population demography and migration [31]. The migration estimates we provide could be used in such a model once reliable estimates of survival and fecundity are obtained. More practically, we demonstrated that there are two distinct harbour seal metapopulations in the UK using genetic data. This confirms previous analyses that showed harbour seals from Southeast UK clustered with samples from France and the Netherlands [18]. In addition, it builds on this work by demonstrating that the previously identified Northwest UK genetic cluster represents two local populations (Northern Ireland and Northwest) and the Northeast UK cluster represents three local populations (Shetland, MFNCO and East Scotland) resulting in five local populations in a northern metapopulation linked by gene flow and dispersal [18]. Furthermore, we have identified putative source–sink dynamics for the northern metapopulation while evaluating if migration patterns had changed between the generations pre- and peri-perturbation. Identifying source–sink metapopulation dynamics and the associated connectivity pattern is of fundamental importance for the management of marine systems [32]. However, this is a particularly difficult task in the case of long-lived vertebrate species. No single type of data (population trends, location tracking, genetics) on its own can achieve this goal. Previous studies have used genetic data indicating asymmetric gene flow to support source–sink dynamics (e.g. [33,34]) but this is not sufficient evidence as net recipients of individuals could still be self-supporting populations. Our integrative approach, combining genetic, location tracking and population trend information, provides a framework for assessing source–sink metapopulation dynamics in future studies. Specifically, we have placed estimates of local population connectivity from genetic and movement data, as well as genetic migrant ancestry information, in the context of population trends to infer whether local populations could be self-supporting sources or immigration-dependent sinks. Through this methodology, we show that the putative key source population of MFNCO has decreased in abundance by perhaps 50% with a concomitant reduction in migration to East Scotland and the Northwest population in the past one generation that spans the start perturbation. Extrinsic factors, such as epizootics, can periodically cause declines and impact pinniped population dynamics [35]. In the case of the Scottish harbour seal, environmental change, including exposure to toxins from harmful algae [36], and competition and predation [17], are hypothesized to be contributing to changes in the population dynamics, but there is no evidence for infectious disease [37]. Future work should examine habitat loss or fragmentation, which previous studies focused on other species suggest can accelerate metapopulation fragmentation and result in regional extinctions [38]. Intrinsic factors, such as density-dependent emigration, have been shown as an important determinant of grey seal metapopulation dynamics [6,39]. If similar mechanisms operate in harbour seals, then the decline in abundance in the MFNCO local population could have led to a concurrent decline in density-dependent emigration to previously connected local populations such as the rapidly declining East Scotland [40]. As pups of the year are thought to be the dispersing age class in harbour seals [41–43], facilitating connectivity across the metapopulation, future work should focus on assessing their patterns of movement and recruitment. According to genetic and pup-tracking data, seals continue to migrate from MFNCO to Shetland peri-perturbation, likely key to the persistence of the Shetland local population. Rather than a regional issue, the decline in the MFNCO local population has had a ripple effect across much of the northern metapopulation. Indeed, the apparent decline in migration from MFNCO to the Northwest local population could have impacts that are yet to be detected or determined. The Northwest local population is also a likely source population for Northern Ireland, as demonstrated by genetic estimates of migration rates and, potentially, pup dispersal. Although the evidence we provide for source–sink dynamics and changes in connectivity are convincing, there are important caveats to consider. Ultimately, our framework uses a range of proxies instead of direct estimates of demographic parameters, such as population trends as an indication that mortality is greater than survival, in the absence of high levels of emigration. As noted earlier, direct estimates of these parameters are needed to definitively assess source–sink and metapopulation dynamics. Furthermore, while we have used genetic data from across the harbour seal's UK distribution to estimate genetic differentiation, migration rates and migrant ancestry, our sample sizes from some locations were small. Future work should assess migration rates using both larger genetic sample sizes and numbers of markers. We also only had pup-tracking data from two of five northern metapopulation local populations. However, our inferences from multiple lines of evidence––genetic, pup and non-pup movement data––were consistent, providing confidence in our results. We hope that this prompts other scientists to examine extant datasets on other species for similar analytical opportunities. Our study uniquely considers population trend, location tracking and genetic data over a multi-generational timescale for a long-lived mammalian species and provides convincing evidence of source–sink metapopulation dynamics for this top predator. The results suggest that the Southeast England SMU can be assessed and managed independently from those in Scotland and Northern Ireland, with implications for the broader management of the species across Europe. Management across the northern metapopulation should consider connectivity patterns identified here. Continued research into habitat preference for UK harbour seals, combined with patterns of connectivity described here and vital rate estimates, will contribute considerably in the near future to the debate on the metapopulation and habitat paradigms for understanding declines of species [44]. More broadly, this work shows that changes in migration and connectivity at a local level can impact wide-scale dynamics, which has important implications for management of the diverse array of terrestrial and marine species that exist as metapopulations. For example, most conservation frameworks assess changes in abundance over time (e.g. the IUCN red list, [45]). This work suggests that changes in migration rates and connectivity could foreshadow changes in abundance. Monitoring and identification of reduced connectivity may prompt conservation measures to be put in place that could forestall decline, or could be assessed retrospectively, as has been done here. As anthropogenic activities cause more widespread environmental degradation and habitat fragmentation [46], understanding connectivity could be an important factor in maintaining both populations and biodiversity in the future. In the UK, the Sea Mammal Research Unit (SMRU) and the University of Aberdeen collected skin samples during live-capture of harbour seals across 20 sampling sites from 2003 to 2012, using methods described in Sharples et al. [16]. In addition to the UK samples, 36 harbour seal samples were included from Norway, Dutch Wadden Sea and France as described in Olsen et al. [18]. DNA was extracted using a salt-saturated technique [47]. Fourteen microsatellite loci were amplified and genotyped by Xelect Ltd (St Andrews, UK). All genotyping previously conducted [18] was repeated to ensure complete comparability across the dataset, but was augmented with more loci and UK samples to increase the power of our analyses (ESM, table S7). Tracking data provide two sources of information: regional movements of individuals aged one year and older within the non-breeding season and movements of pups. We determined movement behaviour in non-pup seals using Argos satellite relay data loggers or GPS/GSM phone tags (developed and supplied by the SMRU Instrumentation Group) deployed from 2001 to 2017 on 334 harbour seals in eight of the SMUs (table 1; see ESM, table S5). We also considered the movements of tagged pups as an indication of dispersal; juveniles are more dispersive than adults in pinnipeds, and the movements of these pups in the first few months of life may be indicative of where they will recruit into the breeding population. However, data were only available from two locations: 46 pups tagged in Orkney, North Coast & Orkeny SMU and West Scotland SMU [42]. The tags (SPOT tags, Wildlife Computers, Redmond, WA, USA) were deployed on flipper tags and thus do not fall off during the annual moult. Pup tag duration was 31–424 days (mean: 155 days) and non-pup tag duration was 28–243 days (mean 95 days). We inferred the metapopulation membership of local populations by estimating genetic differentiation between SMUs, and between sampled haul out sites and subunits within SMUs, using microsatellite data. We calculated pairwise FST values [48] using GENEPOP [49], with significance assessed using the exact G test in the same program (100 000 dememorization steps, 1000 batches each with 10 000 iterations). Furthermore, we investigated isolation-by-distance across the UK by regressing FST/(1-FST) with the log of the ‘at-sea' distance between haul out sites using the ISOLDE program implemented in GENEPOP [50]. To infer recent connectivity, we also ran the program BayesAss ([51], see next section). Finally, we conducted DAPC using the R package adegenet [52] to investigate the genetic differences between SMUs in a multivariate statistical framework. In order to understand migration and genetic connectivity over recent timescales, we used program BayesAss [51]. The program estimates immigration rates over the past two generations using gametic disequilibrium signal generated by immigrant individuals or their descendants (for details see ESM). The patterns of connectivity, in terms of migration rates and ancestry of individuals, were used to infer connectivity over the past two generations. As samples were collected between 2003 and 2012 and the harbour seal is estimated to have a 14.8-year generation span [53], the approximate timings for the migration estimates are 1993 to 2007 (taking midpoint of the samples) for the past one generation, clearly spanning the recent decline (peri-perturbation), and 1977 to 1992 for the second generation, clearly preceding the recent decline (pre-perturbation). We used the tracking data to investigate connectivity through the patterns of movement of UK harbour seals. For both pups and non-pups, we calculated the proportion of animals tagged in each unit that moved between local populations, SMUs or areas within SMUs (north, central and south subunits of West Scotland SMU), identifying movements using haul out locations rather than at sea locations. Of the SMUs which have shown decline, only for East Scotland were there data that could reliably represent both pre- (n = 10/33 tagged in late 2001/early 2002) and peri-perturbation (ESM, table S5). The movements from the Moray Firth represent peri-perturbation (tags deployed from 2004 onwards). For Shetland, all tags were deployed in late 2003/early 2004; the gap in the surveys synonymous with the 40% drop in population size. For North Coast & Orkney, 14/53 tags (14/19 of those deployed in northern Orkney) were deployed during the gap in surveys (late 2003/early 2004), with the remainder tagged peri-perturbation (2011–2017). All tags deployed in Southeast England were deployed after the 2002 PDV epidemic (2003 onwards). We also examined population trend data to describe and categorize the trajectories of the different SMUs as increasing, decreasing, stable or depleted (defined as a decline and then stabilization) using published information [17]. Briefly, harbour seal population trend data was compiled from the 10 SMUs, within which greater than 50 individuals have been counted during a survey. Counts were conducted during the annual moult when the highest proportion of the population is hauled out (c. 0.72%; [54]), ensuring compatibility of data across survey years and regions and as described more thoroughly in [17]. Although the overall trend for UK harbour seals is stable or even increasing abundance, the SMUs exhibit strikingly different dynamics (figure 1). We made inferences about the source–sink dynamics of the harbour seal metapopulations from the genetic and short-term movement estimates in the context of the trends in abundance [17] (figure 1). Specifically, we summarized whether the available data suggested local populations were (i) genetically distinct, based on the estimates of pairwise FST and migration rates, (ii) linked by movement of non-pups and (iii) demographically distinct, based on the proportion of non-migrants from BayesAss and dispersal inferred from the pup-tracking data (where available). Finally, we considered whether these data suggested that the local populations were putative sources or sinks in the context of the trend and abundance data. A local population was considered: (i) putative source region if genetic and demographic (pup movement) data results indicated high internal recruitment and substantial degree of emigration to other regions and non-pup-tracking data indicated low rates of movement or (ii) a putative sink region if genetic and demographic data indicated substantial recruitment from outside the local population and showed a trend similar to its source population. The long-term viability of putative source populations was considered in the context of their population size and trend data [17]. Ultimately, we consider whether the totality of the evidence suggests a broadscale metapopulation decline or a regional decline. All procedures were carried out under Animal (Scientific Procedures) Act, 1986 Home Office Licences issued to SMRU (PIL nos. 60/3303, 60/4009 and 70/7806). Data are available from the Dryad Digital Repository: https://doi.org/10.5061/dryad.j9kd51c8j [55]. E.L.C., O.E.G., A.H., A.O. and D.J.F.R. designed the research; A.O., A.H. and D.J.F.R. performed research; E.L.C., M.T.O., O.E.G. and D.J.F.R. analysed data and all authors wrote the paper. We declare we have no competing interests. Fieldwork and tag deployment were supported by Dept of Business, Energy and Industrial Strategy, Beatrice Offshore Windfarm Ltd, Crown Estate, Highlands & Islands Enterprise, Marine Current Turbines Limited, Marine Scotland Science, Moray Firth Offshore Renewables Limited, Natural Environment Research Council (NERC) and Scottish Natural Heritage (SNH). Genetic analysis was funded by SNH, the Scottish Government and NERC (grant no. SMRU 10/001). Data collection in the Thames was funded by BBC Wildlife Fund and SITA Trust. O.E.G. was supported by the Marine Alliance for Science and Technology for Scotland, funded by the Scottish Funding Council (grant no. HR09011). E.L.C. was supported by a Newton Fellowship (Royal Society of London), Marie Curie Fellowship (EU Horizon2020) and a Rutherford Discovery Fellowship (Royal Society of New Zealand). A.J.H. and D.J.F.R. were supported by NERC (grant no. SMRU 10/001). We thank all the field teams for their hard work in deploying the telemetry tags and collecting samples for genetic analyses, particularly S. Moss; P. Thompson, University of Aberdeen, for the use of the data collected in the Moray Firth; J. Grecian for assisting with analysis. FootnotesElectronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.4979936. References
Page 8A fundamental question in phylogeography and evolutionary and conservation biology is what factors, climatic, geomorphological or other, have influenced the contemporary distribution of genetic diversity and indeed the geographical distribution of species. The glacial cycles of the Quaternary often figure prominently among these factors [1–3]. Our general understanding of the evolutionary influences of these climate cycles has improved dramatically over the last three decades. Initially most phylogeographic studies focused on Northern Hemisphere taxa [4]. More recently, however, a number of studies have examined these questions on Southern Hemisphere species including species from tropical South America [5–8] and its temperate southern cone, Patagonia. In Patagonia, the focus area of the present study, phylogeographic patterns have been described for terrestrial [9–12], coastal marine [13–15]; freshwater [16–23] and diadromous species [24–27]. The processes usually described as the two most important factors in South America are the uplift of the Andes, which began approximately 23 Ma [28], and the glacial cycles of the Quaternary (2.5 Ma–10 000 BP). Their combined influence has affected phylogeographic patterns for many species, including the number and location of glacial refugia and the rates and directions of late and post-Pleistocene gene flow [12,24,29,30]. Some species survived the last glacial maximum (LGM) in glacial refugia east of the Andes, on the largely unglaciated Patagonian steppe [10,12,17,22,31]. A few cold-adapted species also survived in small refugia west of the Andes within the area mostly covered by glaciers (i.e. south of 42° S; [22,32–34]). Refugia have also been identified west of the Andes north of 42° S where the ice cover during the LGM did not reach the Pacific Ocean [16,22,34,35]. As the climate warmed following the LGM and glaciers began to melt, there were significant changes to the Patagonian landscape, especially along the backbone of the Andes. As the glaciers retreated, large proglacial lakes formed in several places along the eastern flank of the Andes [35–41]. Known palaeolakes include Lake Elpalafquen (41° S), Lake Cari Lafquen, on the Patagonian steppe also at approximately 41° S [34, p. 506], Lake Chalenko, which covered present-day lakes Buenos Aires-General Carrera and Pueyrredón-Cochrane [40–41], Lake Caldenius, encompassing present-day lakes Azara, Belgrano, Mogote, Nansen, Volcán and Burmeister (approx. 48° S) [42–44], and Lake Fuegian, on the island of Tierra del Fuego [44]. After formation, and for perhaps several hundred years, the palaeolakes drained to the Atlantic; ice dams at their western limits prevented westerly flow. As melting continued, however, the western ice barriers were breached, sometimes catastrophically, and flow reversed, draining towards the Pacific [38]. Water levels subsided, leaving the current high elevation Patagonian lakes as remnants of the ancient palaeolakes. It has been pointed out previously that phylogeographic patterns for Percicthys trucha and Galaxias platei based chiefly on mitochondrial DNA (mtDNA) polymorphism [16,22], are broadly consistent with post-glacial drainage reversals, but geomorphological evidence for catastrophic drainage reversals is available in detail only for the Baker river valley [38,40]. Here, we revisit the phylogeographic pattern in P. trucha, over its entire geographical range on both sides of the Andes (16 drainages and 46 sampling locations) (figure 1) and using a suite of 53 sequenced nuclear microsatellite DNA markers and individuals (n= 835). The use of a high number of sequenced nuclear microsatellite markers provides two advantages: (i) it addresses the general concern common to most phylogeographic studies, that patterns based on mtDNA polymorphism reflect the coalescence time and evolutionary history of a single gene, and not necessarily that of the organism a whole; and (ii) their relatively high mutation rate resulting in high polymorphism makes microsatellite markers particularly useful for phylogeographic studies, that both focus on relatively short time frames (e.g. Late Pleistocene–Holocene to present) and examine patterns at a finer geographical scale than may often be feasible with mtDNA. We examined the potential consequences that the major climatic cycles of the Quaternary and the resulting changes in regional geomorphology had on the distribution of genetic diversity, and indeed on the geographical distribution of the species. We assess whether the phylogeographic patterns described for P. trucha using mtDNA haplotypes [16] are confirmed by patterns of divergence in genomic DNA, or conversely whether genomic DNA tells a different story regarding the geomorphological history of the region. Our data provide genomic evidence of drainage reversals in several systems for which no detailed geomorphological evidence exists yet, including the Puelo and the Futalaufquen-Yelcho systems in northwestern Patagonia and the Pascua system in southern Patagonia. These results suggest phylogeographic data can complement and indeed assist in the design of geomorphological studies and vice versa while confirming the role that past processes have played in shaping current species distributions. Figure 1. Percichthys trucha sampling locations. West of the Andes south of 42° S, where the LGM reached the edge of the continental shelf, P. trucha is present only in locations associated with trans-Andean systems. Insets: sampling locations in the (a) Valdivia, Bueno and Limay/Negro, (b) trans-Andean Puelo, (c) trans-Andean Futalaufquen/Yelcho, and (d) trans-Andean Baker and Pascua systems. (Online version in colour.)
Percichthys trucha, a freshwater fish species native to Patagonia and neighbouring regions to the north, is widely distributed throughout the region. The species is not harvested nor is it the focus of recreational angling, which in Patagonia focuses on introduced salmonids. The species has not ever been the subject of supplementation with hatchery reared fish in any of the river systems visited for this study. Current phylogeographic and genetic diversity patterns can thus be safely considered to be solely the result of natural, biogeographical and geomorphological processes and drivers. A total of n= 835 Percichthys trucha collected from 16 systems throughout the geographical range of the species in Patagonia (33° S–50° S; electronic supplementary material, table S1; figure 1) were genotyped for 53 microsatellite markers (electronic supplementary material, table S2). Eleven systems drain into the Pacific Ocean; of which six have their headwaters in Chile (Maipo, Nilahue, Rapel, Biobío, Bueno, Maullín), but the other five have their headwaters east of the Andean highest peaks (Valdivia, Puelo, Futalaufquen/Yelcho, Baker, Pascua). The remaining five systems (Colorado, Negro, Chubut, Chalia and Santa Cruz) have their headwaters east of the Andes and drain into the Atlantic (electronic supplementary material, table S1). Sampling was conducted between 1998 and 2008. Lakes were sampled with gillnets (mesh sizes 30–140 mm) and rivers were sampled by electrofishing. Tissue samples (blood or fin clips) were placed in ethanol and subsequently stored at −20°C. Caudal fin clips and blood samples, both approximately 2 mm3, were digested in proteinase K (Bio Basic Inc., Markham, Ontario, Canada) at 55°C for 8 h. A glassmilk protocol modified from [45] was used to extract the DNA using a Perkin Elmer Multiprobe II plus liquid handling system (Perkin Elmer, Waltham, MA, USA). DNA quality and quantity were checked against a standard using gel electrophoresis on 2% agarose gel. Information on primer design and testing is available in the electronic supplementary material. Genotyping was conducted using Megasat [46]. Microsatellite sequence data were checked and the input file was modified as necessary (e.g. changes to flanking regions) to verify that not only were target microsatellite sequences retained, and non-target sequences discarded, but to ensure we were not accidentally discarding useful sequence data, or potentially entire loci. Histogram plots from Megasat were manually verified to ensure correct allele call. Microsatellite loci were removed from further analysis if they did not amplify, were too difficult to score accurately, or were monomorphic. In total, 14 out of 75 loci were removed for these reasons. The remaining 61 were amplified (again using two multiplex reactions, this time consisting of 29 and 32 loci, respectively) and sequenced for the remaining samples. Additional loci filtering steps occurred once data had been collected for all 835 individuals. These steps included removing loci which contained null alleles (identified using Microchecker [47]), exhibited greater than 12% missing data, or were sequenced at low depth (less than 50). BayeScan v. 2.1 [48] was run, using default parameters, to identify putative loci under selection. The final number of loci for all subsequent analyses post filtering was 53 (electronic supplementary material, table S1) with an average of 2.05% missing data per locus. We used the program Structure 2.3.4 [49] to assess population structure hierarchically. We first examined the entire dataset. Identified clusters were then independently subject to further Structure analysis. Structure was run using 500 000 Markov chain Monte Carlo permutations and 100 000 burn-in steps with each K value replicated five times. We used the Evanno method [50] as implemented in Structure Harvester v. 0.6.92 [51] to determine the most likely number of clusters. This process was continued up to the identification of individual rivers or lakes. If needed, geographical location was also taken into consideration. Clumpp 1.1.2 [52] was used to combine the five replicates for the chosen K value into a single output which was then visualized using Distruct 1.1 [53]. Poptree2 [54] was used to create a neighbour-joining phylogenetic tree (with 1000 bootstrap iterations), using Nei's distance (Da), which has a high likelihood of producing a more accurate tree when microsatellite data are used [55]. The phylogenetic tree was then visualized using the Interactive Tree of Life (iTOL) v. 4 [56]. Pairwise FST values were not calculated owing to the high variation in sample sizes (electronic supplementary material, table S1), which can bias estimates [57]. Two analyses of molecular variance (AMOVA) were performed using Arlequin 3.5.2.2 [58]. In the first analysis, we grouped collections by contemporary drainage configuration, with one group comprising all systems currently draining into the Pacific Ocean, regardless of headwater location (i.e. whether headwaters are west or east of the Andes), and the second group comprising all systems currently draining into the Atlantic Ocean. In the second AMOVA, we also considered two groups, but this time all systems with headwaters east of the Andes were grouped together, regardless of whether they currently drain into the Atlantic or Pacific Ocean. This grouping is expected to reflect the drainage scenario during and prior the LGM. An increase in the percentage of among-group variance explained under the second scenario would suggest that P. trucha populations inhabiting trans-Andean systems (those with headwaters east of the Andes but draining into the Pacific Ocean) are more closely related to other Atlantic draining systems than they are to Pacific draining systems. We subsequently repeated this analysis changing the grouping of one trans-Andean system at a time to examine each system's relative influence on patterns of genetic diversity. Arlequin 3.5.2.2 [58] was used also for testing for linkage disequilibrium (LD). We acknowledge that some authors [59] consider that all Percichthys populations in Chile, including those in the southern Chile, are not P. trucha, but a different species entirely, Percichthys chilensis, a view the genetic and phylogeographic data we report in this study do not support (see below). Although the entire dataset consisted of n=835 individuals genotyped at a panel of 53 microsatellite loci, sample sizes per population were relatively small (mean n= 18.2, median n= 20.5). We tested for LD between pairs of loci in the nine collections available with n ≥ 30 (see the electronic supplementary material, table S1). We found no evidence of LD that was significant and consistent across the nine populations tested. We also tested for evidence of selection using all 46 collections, and subsets with relatively large sample sizes. These included the nine collections with n ≥ 30, the four collections within the Limay-Negro system (all n ≥ 26) and the two collections within the Maullín system (both n ≥ 27). While 18, 9, 1 and 0 loci showed up as outliers in each of these tests, respectively, there was no consistency in the identity of the outlier loci across more than two of these four tests. We conclude there is no meaningful evidence of selection in any of these loci and all were retained in all subsequent analyses. The neighbour-joining tree clusters all Chilean populations north of latitude 42° S together and separate from all other populations to the east in Argentina and to the south regardless of whether they drain into the Atlantic or Pacific Oceans (figure 2). Within this group of north-central Chilean collections, collections line up in a pattern that correlates with latitude with the northernmost populations of Maipo and Nilahue (dark blue in figure 2) separated from the remaining populations further to the south. All populations in this cluster, including Maipo and Nilahue, have their headwaters west of the Andes with the exception of Neltume and Panguipulli, which are part of the Valdivia river system, the headwaters of which lie east of the Andes in Lake Lácar in Argentina. At the opposite end of the neighbour-joining tree are all populations with headwaters east of the Andes in Argentina. Some of these populations drain into the Atlantic (e.g. collections from the Colorado, Negro, Chubut, Chalia and Santa Cruz rivers) and some into the Pacific Ocean (e.g. Puelo, Futalaufquen/Yelcho, Baker and Pascua river (Lake San Martín)). Two major clusters are identified in this large group. One cluster comprises all collections from northern Patagonia (from north to south: Colorado and Negro rivers, Puelo (lake and river), and lakes Futalaufquen and Musters), while the second cluster comprises all collections from southern Patagonia (e.g. Baker, Pascua and Santa Cruz river systems). Figure 2. Neighbour-joining tree representing P. trucha from 16 river systems (46 sampling locations: lakes or rivers) throughout the species distribution in Chile and Argentina. At the broadest scale, the tree distinguishes collections from Pacific draining systems north of 42° S from collections with headwaters east of the Andes that drain into the Atlantic and those that are trans-Andean south of 42° S and drain into the Pacific. Within the first group, collections from Maipo and Nilahue (1 and 2) and to a lesser extent, from the Biobío system (3–5) are distinguishable from those to the south (6–14). Within the second group the colouring reflects two glacial refugia east of Andes, one in northern and one in southern Patagonia. Seven collections have sample sizes n ≤ 2 (electronic supplementary material, table S1), regardless, they cluster within their respective river systems. (Online version in colour.)
In the first analysis, when populations were grouped according to present-day drainage (Pacific versus Atlantic) irrespective of whether their headwaters lie east or west of the Andes, the percentage of the total variation explained by the variance among groups is 5% (electronic supplementary material, table S3a). By contrast, this percentage was 26.94% when the collections from the trans-Andean systems: Puelo, Futalaufquen/Yelcho, Baker and Pascua were grouped according to their presumed ancestral drainage direction (electronic supplementary material, table S3b) (24.91% if the collections from lakes Neltume and Panguipulli in the Valdivia river system are pooled with the Atlantic drainage group, see below). We then examined the relative influence of each one of these systems (Puelo, Futalaufquen/Yelcho, Baker and Pascua) on the percentage of variation explained by the variance among groups by grouping them one at a time with the Pacific draining systems. A small decline in the percentage of the variance explained by differences among groups would indicate the system had little influence in the distribution of genetic variance between the Pacific and Atlantic draining groups. In all cases, the percentage of the variance explained by differences among groups dropped considerably from 27% when their ancestral drainage was assumed to have been the Atlantic (electronic supplementary material, table S3b), to 15.45%, 21.75% and 11.23% when grouping each of Puelo, Futalaufquen/Yelcho and Baker individually with all other Pacific draining systems (electronic supplementary material, table S3c–e). The percentage of genetic variation explained by the differences between groups also dropped when pooling the collections from the Pascua River with their contemporary Pacific drainage, but only to 25% (electronic supplementary material, table S3f). This is not surprising given that only 10 individuals were available from a single location (Lake San Martín) within this system. Structure analyses were conducted at four hierarchical levels. At the highest level, all 16 systems organized by drainage and latitude were analysed together. At this level, K = 2 (figure 3a), with the first cluster comprising P. trucha individuals from all systems in Chile north of latitude 42° S. These include the collections from Maipo and Nilahue at the northern edge of the species distribution to the collections from the Maullín system in the south. The second cluster comprises all collections from systems with headwaters east of the Andes from latitude 33° S to 50° S regardless of contemporary drainage (i.e. whether Pacific or Atlantic; these include: Colorado, Negro, Puelo, Futalaufquen/Yelcho, Chubut, Baker, San Martin and Santa Cruz). Notaby, the collections from the northern edge of the species distribution in Chile (Maipo, Nilahue) and in Argentina (Tunuyán, Atuel) appear to exhibit some level of admixture (figure 3a). Figure 3. Hierarchical population structure analyses of P. trucha from 16 drainages. Populations are ordered as in the electronic supplementary material, table S1. (a) All sampling locations, K = 2 distinguishing collections from Pacific draining systems north of latitude 42°S versus Atlantic draining systems throughout the species' distribution with those from trans-Andean systems south of 42° S. (b) Genetic differentiation within each of the groupings found in level (a) above. Two groups are distinguished among the collections from the Atlantic draining systems in Argentina and the trans-Andean systems south of 42° S probably reflecting two glacial refugia east of the Andes, in northern southern Patagonia, respectively. Levels (c) and (d), differences among sampling locations within river systems reflecting influence of genetic drift. Within (c) the Baker system (samples 32–41) is distinguishable from Chalia, Pascua and Santa Cruz systems (samples 42–46). Differences among these sampling locations become apparent in (d). Lake San Martin (43), drains into both the Atlantic and Pacific. (Online version in colour.)
The second hierarchical level (figure 3b) examines separately collections from the two clusters above. At this level, P. trucha from Maipo and Nilahue (populations 1 and 2) appear genetically distinguishable from those from the Rapel and Biobío systems (populations 3–5) and both groups are distinct from collections form further south including those from the Valdivia, Bueno and Maullín systems (populations 6–14), which at this hierarchical level appear indistinguishable from each other (figure 3b). The populations with headwaters in Argentina (populations 16–46) clustered into two groups, again indicating a latitudinal divide with the more northern river systems (Desaguadero, Colorado, Limay and Negro, Puelo, Futalaufquen, and Chubut) grouping together, and a second grouping composed of the more southern systems (Baker, Chalia, Pascua, and Santa Cruz) (figure 3b). At the third hierarchical level (figure 3c), we found: (i) no discernable difference between fish from the Maipo and Nilahue rivers with Nilahue containing some migrant and admixed individuals (figure 3c); (ii) significant differences between collections from the Rapel and Biobío systems (populations 3 versus 4 and 5); and (iii) the remaining populations in this pool (populations 6–14) clustering into two groups and thus requiring a fourth hierarchical level of analysis where differences are uncovered among the Valdivia (populations 6–8), Bueno (populations 9–12) and Maullín systems (figure 3d). We also found differences between and among Atlantic and trans-Andean draining systems to the south as well as among some of the sampling locations (lakes or rivers) within each of these systems (figure 3c). Analyses at this fourth hierarchical level were also conducted with collections from the Baker (populations 32–41) and those from the Chalia, Pascua and Santa Cruz river systems (populations 42–46), where genetic differences are manifested among populations within rivers systems probably largely as a result of genetic drift (figure 3d). In this study, we describe a pattern of genetic divergence among P. trucha populations from throughout the species' range reflecting both the role of the Andes as a geographical barrier to dispersal and that of the glacial cycles of the Quaternary (2.5 Ma–10 000 BP; [35]) in forcing populations to retreat to glacial refugia and in influencing patterns of recolonization. Our analyses, based on a suite of 53 sequenced nuclear microsatellite markers, indicate that P. trucha survived the LGM in at least three glacial refugia. A refugium (or possibly multiple refugia) located west of the Andes and north of where the ice sheets reached the Pacific coast during the LGM (latitude 42° S), served as the source(s) for recolonization of some of the northern river systems west of the Andes. Percichthys trucha also survived the LGM in at least two refugia east of the Andes and east and/or north of the ice sheets, one in north-central Patagonia, and the other in southern Patagonia (figure 3a,b). Atlantic-draining river systems were recolonized from these refugia, and there is a clear distinction between northern systems which were recolonized from the north and/or east, and southern systems which were recolonized from a distinct and probably more southerly refugium. The refugia east of the Andes were also the source of post-glacial colonists for at least five Pacific-draining river systems. These systems all have headwater lakes east of the Andean divide and were probably recolonized during the early stages of deglaciation, when they drained towards the Atlantic. As the glaciers receded, these drainages reversed direction, and currently flow towards the Pacific, but retain P. trucha populations derived from the eastern refugia. In fact, these are the only populations of P. trucha west of the Andes and south of 42° S, indicating (i) the absence of post-LGM north to south migration into this region, and (ii) that unlike Galaxias platei [22] and other more cold-tolerant species [32,34], P. trucha did not persist in small glacial refugia west of the Andes where the ice sheet reached the sea. The deep differentiation between P. trucha populations occupying Atlantic drainages, and those in most Pacific drainages, supports the rise of the Andes (beginning an estimated 23 Ma [28]), as the primary driver of phylogeographic and biogeographic patterns in Patagonian flora and fauna. Since then, phylogeographic patterns have been modified by climatic and geomorphological events, chief of which were the Quaternary glacial cycles [60–61], which caused species ranges to shift, expand and contract, with impacts on genetic diversity as well as distribution and also rearranged Patagonian landscapes and riverscapes ([16,22], this study). The most recent glaciation [60–61] culminated in a 1800 km long ice sheet that covered the Andes from latitude 38° S to 55° S approximately 20 000 BP (figure 1). This ice sheet extended east onto the Patagonian steppe, and west to the Pacific Ocean south of latitude 42° S (figure 1 and [34]). Some species probably survived in local refugia within the glaciated region, while others will have moved north, west or east of the ice sheets. Refugia are known to have existed north and west of the continental ice on the western side of the Andes ([35,62]; figure 1), as well as east and north of the ice sheets on the eastern side ([16] this study). In addition, small refugia appear to have existed within the ice sheet (west of the Andes and south of 42° S) for some species [22,32,33]. Our results demonstrate the power of a relatively large suite of nuclear microsatellite DNA markers to unravel phylogeographic patterns in a widespread freshwater fish, an approach made possible by the recent development of sequence-based protocols and software for microsatellite genotyping [46]. Such protocols provide the advantage that fragments can be sized precisely with many individuals and loci in a single sequencing run, thus reducing genotyping cost and time and minimizing the need for standardization across laboratories [63–65]. Although microsatellites are sequenced, in the present study alleles were scored solely based on the length of the repeated motif using the Megasat [46] platform. Any potential allelic diversity stemming from the existence of single nucleotide polymorphisms (SNPs) within the repeated motifs has thus remained undetected. In the future, the use of sequence-based simple sequence repeats (SSR) that incorporate SNP information within the repeat motif are likely to improve population genetics and phylogenetic analyses and inferences. Regardless, using a combination of a large panel of genetic markers and landscape information reflecting glacial history, we describe phylogeographic patterns for one of the most widespread species of fish in Patagonia, P. trucha. These patterns are consistent with the presence at least two glacial refugia east of the Andes during the LGM, and at least one, and probably more refugia west of the Andes in central Chile (north of 42o S where the ice did not reach the sea). Our study also provides genetic evidence that is consistent with geographically separated and multiple episodes of drainage reversal that influenced the distribution of genetic diversity of P. trucha throughout most of its latitudinal range. We discuss the evidence for an influence of drainage reversal on patterns of genetic diversity and species distribution in more detail below. Populations of P. trucha currently inhabiting river systems with their headwaters west of the Andean divide are all found north of 42° S, where the ice sheet did not reach the Pacific coast during the LGM [35]. These populations either persisted in local refugia (possibly downstream within the same river system) or these systems must have been recolonized from the north. The four northernmost populations, those from the rivers Maipo and Nilahue (samples 1 and 2 in the electronic supplementary material, table S1), and those from the Rapel and Valdivia systems (samples 3–5) were genetically distinguishable from the other Pacific draining systems (figures 2 and 3), suggesting that these four populations probably survived in refugia separate from the rest. The relative geographical isolation of these river systems, particularly that of the Maipo and Nilahue, has probably helped the populations maintain their distinctiveness (figure 1). Among the other Pacific draining systems with headwaters west of the Andes, we find less genetic differentiation than for the northern four systems but more than is seen among the Atlantic draining systems (note branch lengths in figure 2). In addition, there is a shallow signal of differentiation that follows a latitudinal gradient from north to south. We infer, therefore, that the populations in these Pacific draining systems probably persisted through the LGM in deglaciated coastal areas, but that they were also influenced by predominantly north to south gene flow. North to south recolonization on the western side of the Andes has been observed in terrestrial species, such as the forest-dwelling mouse Abrothrix olivaceus and the steppe-dwelling mouse Abrothrix xanthorhinus [66]. Gene flow among aquatic populations occupying different watersheds may have occurred during periods of high discharge as the ice sheet melted. River systems with headwaters east of the Andes that harbour P. trucha comprise all major river systems that currently drain into the Atlantic Ocean (the Colorado, Negro, Chubut and Santa Cruz river systems) and several Pacific-draining, trans-Andean systems (the Puelo, Futalaufquen/Yelcho, Baker and Pascua river systems) (figure 1; electronic supplementary material, table S1). These populations formed two genetically differentiated groups, suggesting that they were colonized from two separate glacial refugia. The P. trucha populations currently found in the Desaguadero, Colorado, Negro and Puelo river systems form a group, and probably originated from a northwestern refugium, while populations in the Chubut, Santa Cruz, Futalaufquen/Yelcho, Baker, Pascua and Chalia catchments probably originate from a southern refugium. The presence of southern refugia has also been proposed for some terrestrial taxa [12,31,67]. Southern refugia were probably facilitated by the greatly expanded Patagonian steppe area east of the ice sheet, with much of the continental shelf exposed as sea-level fell during glacial periods [68,69], and the persistence of such refugia may have been enhanced by the high connectivity among drainages afforded by the braided and deltaic connections that formed and re-formed on the shelf during glacial periods [69,70]. Thus, the northern and southern genetic groupings both contain populations of P. trucha from present-day Atlantic drainages and populations in trans-Andean systems, where the headwaters lie east of the Andes, but the system drains west to the Pacific (figure 3b,c). There is strong evidence from geomorphological studies of the Baker [38,40–41] trans-Andean system for drainage reversal following the LGM. At the LGM, the lakes at the headwaters of the Baker system were part of a large eastward draining proglacial lake (Lake Chalenko) that extended from approximately latitude 46° S to 48° S, with raised deltas tens to hundred of metres above contemporary valley floors still visible [36,37,39,71]. As the glaciers melted, the ice dam that had formed the western limit of the lake was breeched (ca 12 000 BP [38]), and the water began flowing west, through valleys previously blocked by ice ([35,38–40] and references therein). The genetic similarity of P. trucha populations in the Baker system with populations in Atlantic-draining systems at similar latitudes is thus consistent with the geomorphological findings. In this study, we have identified three additional river systems that have probably undergone post-LGM drainage reversals: the Puelo, Futalaufquen/Yelcho, and the Pascua systems. All are trans-Andean systems, Pacific drainages with headwaters east of the Andes, and all contain populations of P. trucha that resemble populations from Atlantic drainages at similar latitudes. It is known that proglacial lakes extended over broad latitudinal ranges (42–49° S and 51–53° S) [38,40,41,71–74], but to our knowledge, detailed geomorphological studies of the type conducted for the Baker system have not been done for the other three. Percichthys trucha inhabiting the trans-Andean Valdivia river system (latitude approx. 40° S, headwaters in Lake Lácar in Argentina) comprise an exception to the pattern of higher genetic similarity with Atlantic than with Pacific draining systems described above for the other four southern trans-Andean systems (Puelo, Futalaufquen/Yelcho, Baker, and Pascua). Grouping the collections from the Valdivia system (i.e. Panguipulli and Neltume) with those of Atlantic drainage decreased the percentage of genetic variation explained by groups (from approx. 27% to 25%) suggesting that P. trucha inhabiting these lakes originate from ancestral populations that survived in local refugia, probably preventing the dispersal and expansion of P. trucha from east of the Andean highest peaks by a ‘Founder takes all' effect [75]. South of latitude 42° S, where the ice sheet reached the Pacific during the LGM, there are no P. trucha populations outside of trans-Andean river systems. Drainage reversals are thus the only mechanism by which P. trucha was able to colonize this region after the glaciers receded, and there appears to have been no gene flow between drainages since that time. Drainage reversals are thus responsible, not only for patterns of genetic diversity, but also for the current geographical range of the species. This pattern of distribution is shared with at least one other native Patagonian fish, a catfish of the family Diplomystidae [30], providing further evidence of the importance for the biogeographic patterns of aquatic taxa in Patagonia of the late Pleistocene glacial melt induced drainage reversals. In summary, we have presented evidence based on a suite of 53 sequenced nuclear microsatellite DNA markers that, after the LGM, P. trucha, a widespread Patagonian fish, colonized the western side of the Andes from the eastern side, and that this occurred in at least four distinct river systems, the Puelo, the Futalaufquen/Yelcho, the Baker, and the Pascua systems. While detailed geomorphological evidence exists only for the Baker system [38], drainage reversals during the retreat of the glaciers appear to be the most likely mechanism for all. In addition, we describe evidence for at least three glacial refugia for P. trucha, one (possibly more) west of the Andes north of 42° S, and the other two, east of the Andes, the first in northwestern and northern Patagonia and the second on the southern Patagonian steppe. Conversely, we find no evidence that P. trucha survived the LGM west of the Andes south of 42° S. The presence of P. trucha populations west of the Andes in these southern regions is a consequence of dispersal from the east, most likely owing to drainage reversals that took place in several river systems following the LGM. Our study, therefore, highlights the synergistic value of combining genetic data from a large panel of sequenced microsatellite DNA with information on landscape evolution to develop and test biogeographic hypotheses. Field collections in both Argentina and Chile were conducted under national, provincial and National Parks permits as appropriate. Microsatellite genotypes are available in the Dryad Digital Repository (https://doi.org/10.5061/dryad.n8pk0p2s5) [76]. D.E.R., E.H. and S.J.W. participated in the fieldwork. G.R.M. developed and tested the 75 microsatellite markers. A.P.S. and G.R.M. conducted the molecular laboratory work. D.E.R., A.P.S. and G.R.M. performed analysis. D.E.R. wrote the paper with input from all authors. Authors declare no competing interests. We thank the Committee for Research and Exploration of the National Geographic Society, Washington, for generous support for fieldwork in 2001 (NGS 6799-00) and 2007 (NGS 8168-07), and NSERC Discovery grants and a Special Research Opportunities award (SROPJ/326493-06), as well as Universidad de Concepción (DIUC-Patagonia 205.310.042-ISP) and FONDECYT (no. 1080082) grants which are gratefully acknowledged. Some samples were collected with support from a 2006 to 2010 NSF-PIRE award (OISE 0530267). We thank the numerous colleagues from Universidad de Concepción, Universidad Nacional del Comahue (Argentina) and Dalhousie University who assisted with sample collections during more than two decades of fieldwork in Patagonia. FootnotesElectronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.4979912. © 2020 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original author and source are credited. References
Page 9The ability to discriminate between kin and non-kin promotes the evolution and maintenance of sociality because it allows altruistic behaviours to be directed towards related individuals [1]. Nest-mate recognition, the process by which social insects recognize individuals belonging to their colony, is encoded by cuticular hydrocarbons, apolar lipids found on the cuticle of the majority of insect taxa that primarily function as a waxy barrier that prevents desiccation [2–6]. Within a social insect species, individuals have qualitatively similar hydrocarbon profiles, but the profile differs in relative proportions depending on the colony of origin. Cuticular hydrocarbons are homogenized throughout a colony via trophallaxis (the exchange of regurgitated liquid) and allogrooming between colony members [6–9], creating a gestalt colony odour [10]. In addition to nest-mate recognition and desiccation prevention, social insects also rely on hydrocarbons to encode information about an individual's reproductive and dominance status and task within the colony [11–13]. Despite the central role of cuticular hydrocarbons in insect societies, we still know very little about fundamental genetic and evolutionary features shaping them, including the relative contribution of genetic and environmental factors to phenotypic variation in hydrocarbon profiles within and between colonies [14,15], and how natural selection acts on this variation. Numerous studies have demonstrated that the social insect hydrocarbon profile is influenced by genotype, by tracking familial lines [7,16–18], using cross-fostering designs [8], or demonstrating an association between hydrocarbon diversity and within-colony genetic variation [19,20]. However, very few studies have examined the underlying genetic architecture of social insect hydrocarbons within a formal quantitative genetic framework [21]. Traditional quantitative genetic crossing and pedigree-based mapping populations provide a powerful means to elucidate the contribution of genetic and environmental factors to variation in hydrocarbon profiles [22–26]. The genetic architecture of social insect cuticular hydrocarbons is expected to be more complex than solitary insect cuticular hydrocarbons because it depends on the collective genetic makeup of the colony [27,28]. In social insects, the hydrocarbon profile of each individual can be made up of compounds synthesized directly by the individual itself, as well as compounds synthesized by nest-mates and socially transferred to the individual [6–9,28]. More generally, in social organisms such as social insects, an individual's own traits can be influenced directly by its own genotype (i.e. direct genetic effects) but also indirectly via the genotype of social partners (i.e. indirect genetic effects; [27,28]). To fully understand the hydrocarbon profile's potential evolutionary response to natural selection, in addition to understanding quantitative genetic parameters such as heritability and genetic correlations, we must also understand the fitness consequences of phenotypic variation in the hydrocarbon profile. Knowledge of the fitness consequences of trait variation is necessary to characterize the type (e.g. directional, stabilizing or disruptive) and strength of natural selection acting on a trait [29–32]. Variation in the social insect hydrocarbon profile may affect individual survival and colony productivity by affecting desiccation resistance [33–36], or by influencing chemical communication among nest-mates and the collective behaviour of the colony [11–13,34]. Hydrocarbon structural classes (i.e. alkenes, linear alkanes, monomethyl alkanes and dimethyl alkanes) have distinct functional properties that are likely to influence the roles they play in insect societies and how they are shaped by natural selection [14,20,37,38]. Linear alkanes provide the best desiccation resistance because these molecules tightly aggregate [37,39]. On the other hand, alkenes and monomethyl and dimethyl alkanes are expected to play a larger role in chemical communication because they can be distinguished by the position of their double bond or of the methyl group(s), while linear alkanes can only be distinguished based on their chain length [40,41]. This increased complexity allows alkenes and monomethyl and dimethyl alkanes to encode more information. There is evidence that linear alkanes are less heritable and not transferred between workers as much as monomethyl and dimethyl alkanes, suggesting that linear alkanes are less informative for nest-mate recognition [8]. Here, we use a genetically highly variable laboratory population of pharaoh ant (Monomorium pharaonis) [42,43]. We extracted hydrocarbons from pools of workers and used the pedigree information of the colonies to estimate the heritability of and the genetic correlations between hydrocarbons. Additionally, we used a random forest analysis to identify hydrocarbons that best discriminate between our M. pharaonis colony genotypes. Finally, we estimated the strength and pattern of natural selection putatively acting on hydrocarbons in the laboratory. Monomorium pharaonis colonies primarily live in association with humans both in the tropics in their presumed native range and in introduced temperate regions [44]. Colonies contain multiple queens and produce new colonies by budding [45,46]. We used 48 laboratory-reared M. pharaonis colonies (hereafter ‘colony genotypes’) of known pedigree from our heterogeneous stock mapping population, which was created from eight initial laboratory stock colonies that were systematically intercrossed for nine generations (electronic supplementary material, figures S1 and S2; see [42,43] for details). We split each colony genotype into three equally sized replicates (hereafter ‘colony replicates’) that initially consisted of four queens, 400 ± 40 workers, 60 ± 6 eggs, 50 ± 5 first instar larvae, 20 ± 2 second instar larvae, 70 ± 7 third instar larvae, 20 ± 2 prepupae and 60 ± 6 worker pupae (electronic supplementary material, figure S3). These colony demographics represent a typical distribution found in a relatively small M. pharaonis colony [46,47]. We set up these colonies in separate blocks (usually consisting of 15–18 replicate colonies) between May and November 2016. We maintained all colony replicates on a 12 : 12 h light : dark cycle and at 27 ± 1°C and 50% relative humidity. We fed each colony replicate twice per week with an agar-based synthetic diet [48] and mealworms. Water was provided ad libitum via a glass tube plugged with cotton. Colony replicates nested between two glass slides (5 cm × 10 cm) housed in a plastic container (18.5 cm × 10.5 cm × 10.5 cm) lined with FluonⓇ. As reported in [43], we surveyed each colony replicate for five collective behaviours and two measures of colony productivity (electronic supplementary material, figure S3): (i) foraging, (ii) aggression, (iii) exploratory rate, (iv) group exploration, and (v) colony exploration (see [43] for details). After we completed the behavioural assays, the queens from each colony replicate were removed to trigger the production of new queens and males [49,50]. We conducted weekly surveys of the number of worker, gyne (i.e. virgin queens) and male pupae produced, until all brood matured. We quantified: (i) the total number of sexual pupae (i.e. gynes plus males), and (ii) the total number of worker pupae as measures of colony productivity. Many of the collective behaviours were phenotypically correlated with each other and foraging and exploratory rate were associated with colony productivity (see [43] for full details). Upon completion of behavioural and productivity surveys, from each colony replicate, we collected three samples, each consisting of 15 workers, for the extraction and analysis of cuticular hydrocarbons (electronic supplementary material, figure S3). To extract the mean cuticular hydrocarbon profile of each replicate colony, we transferred each group of workers into a clean 2 ml glass vial and rinsed in 200 µl of high performance liquid chromatography grade (99%) pentane for 10 min. We injected 3 µl of each extract into an Agilent 6890 N gas chromatograph (GC) coupled with a 5375 Agilent mass spectrometer (MS). We identified 34 chemical compounds (figure 1a) by their retention times, fragmentation patterns and comparison with published results [51,52]. We integrated the area under each peak using MSD Chemstation. As they had similar retention times, some compounds co-eluted into the same peak (peak 2: y-C25:1 with peak 3: n-C25; peak 27: x-C31:1 with peak 28: y-C31:1; figure 1a). We combined the areas of each co-eluting pair, leaving 32 peaks available for statistical analysis. For full details, see the electronic supplementary material, file S1. Figure 1. (a) GC-MS spectrum of M. pharaonis cuticular hydrocarbons. All 34 identified peaks are numbered and colour-coded by structural class. Unidentified peaks were either contaminants or unidentifiable compounds. (b) Caterpillar plot showing heritability estimates of individual hydrocarbon compounds, with associated 95% confidence intervals, obtained from univariate animal models. Peak numbers shown in (a) are reported below the corresponding compound(s). Compounds in the plot are grouped and colour-coded by structural class and ordered by chain length (linear alkanes and alkenes) or by chain length and methyl position (mono- and dimethyl alkanes). (Online version in colour.)
We discarded 175 out of 432 cuticular hydrocarbon samples due to contamination and other technical failures, leaving 257 samples from 111 colony replicates (48 colony genotypes) for statistical analysis. The number of used cuticular hydrocarbon samples per colony replicate is reported in the electronic supplementary material, table S1. We standardized the raw hydrocarbon peak areas of each sample using the log-ratio transformation [53], as has commonly been used in the analysis of hydrocarbon data. As in our dataset, some samples had zero area values for certain peaks (electronic supplementary material, table S1), we added a small constant value (0.001) to each peak prior applying the transformation [54,55]. We also separately used a multiplicative simple replacement method [56] to deal with the zeros to verify our results (see Results section). For each colony replicate, we had a single measure for each of the five collective behaviours and for colony productivity but as many as three measures for hydrocarbon peak values. Therefore, we used the mean hydrocarbon replicate value when estimating phenotypic correlations and selection gradients (see below). We performed all statistical analyses in R v. 3.4.1 [57]. A detailed R markdown file including the R scripts, as well as a detailed explanation of each analysis, is included as electronic supplementary material, file S1. To assess narrow sense heritability (h2; defined from 0 to 1) and genetic correlations (rG; defined from −1 to 1) of cuticular hydrocarbon compounds, we analysed our data using the ‘animal model’ approach [58] with the R package ‘MCMCglmm’ [59]. This mixed-effect model uses a Bayesian Markov chain Monte Carlo approach to decompose phenotypic variance into its genetic and environmental components, allowing the estimation of quantitative genetic parameters. In an animal model, the pedigree of many individuals is used to make inferences about expected patterns of genetic relatedness, and together with observed patterns of phenotypic resemblance among individuals, heritability for measured traits and genetic correlations between traits are estimated. We treated our replicate colonies as ‘individuals’ in an animal model, and the pedigree of each colony traced the parents of the worker offspring through the mapping population. While the hydrocarbon profile of each individual worker is expected to depend both directly on its own genotype (direct genetic effects) and indirectly on the genotypes of its nest-mates (indirect genetic effects; [8,27]), our approach does not enable us to estimate the separate contributions of these direct and indirect genetic effects to the observed composite hydrocarbon profile of our replicate worker groups. That is, we cannot quantify the degree to which the hydrocarbon profile of each individual depends on compounds synthesized by that individual, as opposed to compounds synthesized by social partners, but we can quantify the degree to which phenotypic variation in the hydrocarbon profile of groups of workers is predicted by the genotypic makeup of those workers. Similarly, previous animal breeding studies have shown that the total contribution of direct and indirect genetic effects to total genetic variance and total heritability can be estimated by quantifying phenotypic variation among groups of individuals (e.g. [60]), although it is not possible to empirically tease apart the separate contribution of direct and indirect genetic effects. We ran Bayesian univariate models to estimate the heritability of each hydrocarbon compound, and bivariate models (one for each pairwise combination of hydrocarbon variables) to calculate genetic correlations between compounds. Additionally, to verify the results of our univariate heritability models, we ran bivariate models for all combinations of hydrocarbons and took the average of these to get an additional heritability estimate for each hydrocarbon. Univariate and bivariate models had the same random and fixed effect structure, and differed only in terms of priors specification (electronic supplementary material, file S1). We included in the models individual identity, environmental variance and block (samples were collected at different time points) as random factors. Details of model specification are described in the electronic supplementary material, file S1. The total number of new reproductives (gynes and males) produced by colonies is a natural measure of colony-level investment in reproduction, and hence a measure of colony-level fitness [61]. However, because M. pharaonis queens cannot form new colonies without workers (i.e. colonies reproduce by budding; [45,46]), we also quantified a second measure of colony-level fitness, the total number of new workers produced. We estimated strength and type of selection (e.g. directional, stabilizing or disruptive) acting on individual hydrocarbons using a regression approach [32]. Briefly, we first estimated the fitness function relating colony productivity to the abundance of a specific hydrocarbon with a generalized additive model (GAM), using the R package ‘mgcv’ [62]. Then, we obtained linear (β) and quadratic (γ) selection gradients from the fitted GAM model using the function gam.gradients in the package ‘gsg’ [63]. Prior to running the model, hydrocarbon variables were mean-centred and variance standardized. Details of model specification are described in the electronic supplementary material, file S1. We adjusted p-values using the false discovery rate (FDR) method. Our heritability and selection gradient estimates described above are univariate, rather than multivariate, because mixed models can experience issues with model convergence when a large number of traits are included in one model [58,64]. To also consider multivariate models, we first conducted a principal components analysis (PCA) to reduce the dimensionality of our dataset. We included 29 of the 32 hydrocarbons in the PCA, excluding the three hydrocarbons (peaks 23, 26 and 31) that had zero values in the data. We excluded these peaks because PCA is very sensitive to small values, and samples with zeros were clear outliers in a PCA including them (results not shown). We kept the first eight principal components (PCs) which explained approximately 90% of the variation in our dataset (electronic supplementary material, figures S4–S7 and table S2; see the electronic supplementary material, table S3 for PC loadings). We subsequently used the same approaches described above to estimate the heritability and selection gradients of the eight PCs, but included all eight PCs in all models. We did not estimate the genetic correlations between the PCs because, by definition, PCs are orthogonal to each other and, therefore, unlikely to be correlated. We ran Spearman's rank-order correlations to evaluate the strength and direction of association between cuticular hydrocarbons and collective behaviours. We ran a model between each compound and each of the five collective behaviours (160 models in total; [43]), with adjusted p-values using the FDR method. Finally, we used a random forest (RF) classification analysis [65] to determine which cuticular hydrocarbon peaks can best discriminate across the 48 colony genotypes. Although this method does not take into account pedigree relationships, it can provide hints about which hydrocarbons are more variable among colony genotypes, thus highlighting compounds that might be involved in nest-mate recognition. We ran a stratified sampling RF classification model with replacement using the R package ‘randomForest’ [66], and we considered hydrocarbon samples from colony replicates belonging to the same colony genotype as part of one of the 48 colony genotypes classes. We used the mean decrease in model accuracy [67] to interpret hydrocarbons importance in classifying the colony genotypes. We tested whether hydrocarbon structural classes varied in their ability to discriminate between colony genotypes using a linear model. Model details and specifications are included in the electronic supplementary material, file S1. We estimated the heritability of individual hydrocarbons to be between 0.006 and 0.36, with a median estimated heritability of 0.17, in our univariate models (figure 1b). The mean of our bivariate heritability estimates was very similar to our univariate estimates (electronic supplementary material, figure S8). We estimated the heritability of the eight PCs to be between 0.004 and 0.20, with a median estimated heritability of 0.15 (electronic supplementary material, figure S9). Our pairwise genetic correlations estimates indicate that strong genetic correlations are common as 203 out of 503 (40.3%) estimates were greater than 0.2 or less than −0.2 and most of these (167) were positive (figure 2). Strong, positive genetic correlations were especially common between two linear alkanes or between two alkenes (figure 2). Figure 2. Heatmap showing genetic correlation estimates among cuticular hydrocarbons obtained from bivariate animal models. Compounds are grouped by structural class and ordered by chain length (linear alkanes and alkenes) or by chain length and methyl position (mono- and dimethyl alkanes). Different colours indicate the magnitude and direction of the correlation. (Online version in colour.)
To ensure the additions of a small constant value did not skew our results, we also tried a multiplicative simple replacement method [56] to deal with the zeros in our data. Using this approach, we re-estimated the heritability of and genetic correlations between a subset of our hydrocarbons, focusing on the three peaks (23, 26 and 31) that contained zeros. The heritability and genetic correlation estimates were extremely similar between the two methods of dealing with zeros (electronic supplementary material, tables S4 and S5). We report productivity data for each replicate colony in the electronic supplementary material, table S1. Our two definitions of fitness (the production of reproductives and workers) were positively correlated (Spearman's rank, ρ = 0.611, p < 0.001). We found evidence for significant positive or negative linear selection for 10 and 6 hydrocarbons when defining fitness as the production of reproductives or workers, respectively (figure 3; see the electronic supplementary material, table S6 for estimates, s.e. and p-values). All quadratic selection estimates were not significant. Additionally, we found evidence for positive linear selection on the first PC when defining fitness as the production of workers (see the electronic supplementary material, table S7 for estimates, non-parametric case-bootstrapped s.e. and p-values). Figure 3. (a) Caterpillar plot showing linear selection estimates, with associated case-bootstrapped standard errors, for individual hydrocarbons using sexual pupae production as a measure of colony fitness. Compounds in the plot are grouped and colour-coded by structural class and ordered by chain length (linear alkanes and alkenes), or by chain length and methyl position (mono- and dimethyl alkanes). Compounds showing a statistically significant selection gradient are labelled in bold. (b,c) Representative fitness landscapes for sexual pupae production as a function of population mean phenotype values of y-C29:1 and n-C26, respectively. (Online version in colour.)
We report behavioural data for each replicate colony in the electronic supplementary material, table S1. We found significant phenotypic correlations between individual hydrocarbons and four of the five collective behaviours (foraging, aggression, colony exploration and group exploration) (electronic supplementary material, figure S10). We found that 20 of the 32 hydrocarbons were correlated with foraging, nine with group exploration, two with aggression and one with colony exploration. The error rate of the RF model was 17.1%, indicating that most of the hydrocarbon samples were assigned to the correct colony genotype (electronic supplementary material, file S1 and table S8). Two compounds, 11,15-diMeC27 and 7-MeC27, showed the best discrimination accuracy (electronic supplementary material, figure S11). Alkenes (t = 2.743, p = 0.011) and monomethyl alkanes (t = 2.261, p = 0.032) showed higher discrimination power than linear alkanes. There were no other differences in discrimination accuracy among pairwise combinations of structural classes. In solitary insects, many studies have characterized the heritability (e.g. [22,24]) and patterns of natural and sexual selection shaping the cuticular hydrocarbon profile (e.g. [26,68,69]). By contrast, little is known about the heritability and patterns of natural selection shaping the cuticular hydrocarbon profiles in social insects, despite the well-known role that cuticular hydrocarbon profiles play in nest-mate recognition, intra-colony signalling of task and reproductive and dominance status [8,13,21]. Here, we begin to elucidate the genetic architecture underlying variation in the hydrocarbon profile and to characterize how selection acts on it in a laboratory population of the ant M. pharaonis. We provide evidence that the hydrocarbon profile is heritable, shaped by selection and that many hydrocarbons, especially linear alkanes and alkenes, are genetically correlated with each other. We estimated the total heritability of individual hydrocarbons to be between 0.006 and 0.36 with a median of 0.17 (figure 1b). Our heritability estimates of PCs were very similar, between 0.004 and 0.20 with a median of 0.14. These estimates are broadly similar to the estimated heritability for collective behaviour body size, caste ratio and sex ratio made with the same population [43], and are also similar to the range of heritability for individual hydrocarbons estimated from solitary and gregarious insect populations (e.g. [22,24,64]). We expected that compounds with high heritability estimates would also be among the best at distinguishing between colony genotypes in the RF analysis. In accordance with this prediction, the two compounds with the highest heritability estimates (11,15-diMeC27 and 7-MeC27) were also two of the top three compounds at distinguishing between colony genotypes in the random forest analysis. Many of the compounds with relatively high heritability in our study were also highly variable in a previous study of variation in hydrocarbon profiles among 36 M. pharaonis colonies collected at sites around the world [70]. Our study, together with this previous study, indicates that heritable variation for many cuticular hydrocarbons is maintained in M. pharaonis, both in our laboratory mapping population and in nature. As described above, cuticular hydrocarbons play key roles in nest-mate recognition, and hence mediate aggression between colonies. However, the degree to which the necessary variation for genetically based recognition cues is maintained within populations remains broadly unclear [71]. Our results indicate that most compounds making up the cuticular hydrocarbon profile harbour genetic variation that could be informative for genetically based nest-mate recognition or mate choice. Alkenes and monomethyl and dimethyl alkanes are expected to play a larger role in chemical communication (e.g. nest-mate recognition) than linear alkanes because they can be distinguished by the position of the double bond or methyl group(s), while linear alkanes can only be distinguished based on their chain length [40,41]. In support of this prediction, previous work in ants found that monomethyl alkanes were more heritable than linear alkanes, suggesting that monomethyl alkanes are better indicators of colony membership [8]. Our results mostly support this prediction as well. For example, our RF analysis revealed that alkenes and monomethyl alkanes had a higher discrimination power than linear alkanes (electronic supplementary material, figure S11). The genetic correlation estimates between many pairs of hydrocarbons were high, in particular between pairs of linear alkanes or alkenes (figure 2). Similarly, previous studies in fruit flies found many strong positive genetic correlations between individual hydrocarbons [24,72]. Such genetic correlations between hydrocarbons are not surprising, especially between compounds of the same structural class, because the production of different hydrocarbons involves many of the same biosynthetic processes (reviewed by Ginzel & Blomquist [73]). Overall, these genetic correlations mean that the independent evolution of hydrocarbons will be constrained to some degree. Our study is, to our knowledge, the first social insect study to link variation in cuticular hydrocarbons with variation in colony productivity, although previous social insect studies have linked variation in cuticular hydrocarbons to worker survival [35] or to climatic or biotic variation [14,20,34,35]. We defined fitness in two ways: as the production of either new reproductives or new workers. Interestingly, we found similar linear selection patterns using both definitions, as all significant linear estimates were in the same direction between the two fitness measures (figure 3; electronic supplementary material, table S6). This suggests that the hydrocarbon profile optima are largely aligned for the production of both reproductives and workers in our study population. We note that a study of natural selection in a natural population of the red harvester ant Pogonomyrmex barbatus found no relationship between the number of gynes a colony produced and the number of its daughter colonies that survived at least 1 year [74], calling into question whether the number of reproductives produced by a colony is actually a good measure of fitness. However, this is likely in part because of very high mortality by queens attempting to found colonies independently [75,76], and this high variation in the mortality of new queens can be considered a component of offspring (i.e. new queen) fitness that depends on new queen traits, and not parent fitness that depends on the traits of the parental colony (see [77]). Moreover, we suggest that considering both the number of new reproductives and the number of new workers produced are probably better estimates of colony-level fitness in a species like M. pharaonis that reproduces by budding, where new queens do not found new colonies independently, but are accompanied by multiple other queens and workers. These selection results beg the question: what is the likely causal link between variation in worker cuticular hydrocarbons and variation in colony productivity in our laboratory study population? Interestingly, in our study population, the relative abundance of many cuticular hydrocarbons was phenotypically correlated with the collective behaviour of replicate colonies (electronic supplementary material, figure S10), in particular the foraging rate, which in turn was also positively associated with colony productivity [43]. This relationship between variation in cuticular hydrocarbon profile, foraging rate and colony productivity could be mediated by effects of hydrocarbons on the desiccation resistance of workers [33,36]. However, our colonies probably experienced relatively low water stress because the colonies were kept in climate-controlled chambers at 50% humidity, and the colonies always had access to water. Alternatively, we speculate that worker hydrocarbon profiles might influence colony-level division of labour or task allocation [11,12,36,78,79], which could in turn influence foraging rate and colony productivity. As described above, in addition to effects on desiccation resistance, ant cuticular hydrocarbons are well known to influence nest-mate recognition and inter-colonial aggression in many ant species, including M. pharaonis [80]. However, colonies in our study were isolated from each other throughout the course of the study, so that differences in the outcome of aggressive encounters between colonies that probably contribute to differences in nature for colony survival and productivity [81] cannot explain the patterns of selection on hydrocarbon profile that we observed in our laboratory population. Similarly, while cuticular hydrocarbon profiles might also mediate mate choice in M. pharaonis, such a mechanism could not explain the association between worker hydrocarbons and colony productivity that we observed. An interesting complication of the genetic architecture of social insect cuticular hydrocarbon profiles is that the social environment experienced by each individual within a social insect colony strongly influences its hydrocarbon profile, because hydrocarbons are mixed throughout the colony via trophallaxis and allogrooming between colony members [4,6–8]. As a result, the genetic architecture of the hydrocarbon profile, like other socially influenced traits, depends on the collective genetic makeup of colonies [27,28]. Because we quantified the cuticular hydrocarbon profile of groups of workers from each colony, we were not able to distinguish between hydrocarbons that were readily transferred among nest-mates and those that were only produced by a subset of workers and not transferred (see [8]), or to separately estimate the contribution of variation in direct versus indirect genetic effects to estimated total heritability [60,82]. Furthermore, we were not able to consider differences in cuticular hydrocarbons between individual workers based on age, task within the colony, or differences in genotypes within a colony. We conducted the current study in a laboratory environment, which enabled us to strictly control the colony demography (i.e. queen number, worker number, etc.), diet and environmental conditions experienced by the colonies. Such control in particular is valuable, given the complexity of social insect colonies [27,83] and the sensitivity of the hydrocarbon profile to changes in the environment or diet [7,14,15,84,85]. Because M. pharaonis tends to be found in association with humans, we speculate that the laboratory conditions of our study might be more similar to the natural conditions experienced by M. pharaonis, when compared with other non-synanthropic species. Although future field studies are necessary, in particular to determine how variation in cuticular hydrocarbons affects colony productivity in a natural setting and whether the patterns of selection we observed in the laboratory are consistent in nature, a field study on a similar scale as our study is probably not feasible. Overall, this study increases our understanding of the genetic architecture of the hydrocarbon profile and demonstrates that the hydrocarbon profile is shaped by natural selection. Although numerous genes underlying variation in the hydrocarbon profile have been identified in Drosophila [25], the hydrocarbon profile performs different functions in social insects and, therefore, future studies should focus on determining whether the same genes are involved in the expression of social insect hydrocarbon profiles, and how variation in these genes affects variation in the hydrocarbon profile. For example, a recent study used a candidate gene approach and found that inotocin, a peptide similar to oxytocin/vasopressin, regulates the production of hydrocarbons in the ant Camponotus fellah [86]. Future studies should use unbiased approaches such as quantitative trait locus mapping in mapping populations (e.g. [87]) such as ours, and association mapping in natural populations (e.g. [88]). The data supporting this paper are included in the electronic supplementary material files. J.W., L.P., P.d.E. and T.A.L.: conceptualization and methodology; J.W., L.P., T.A.L. and P.d.E.: experimental design and writing—review and editing; L.P. and J.W.: data analysis and writing—original draft; P.d.E. and T.A.L.: project supervision; T.A.L.: funding acquisition. We have no competing interests. This work was supported by National Science Foundation grant no. IOS-1452520 awarded to T.A.L. We thank Chloé Leroy for conducting the GC injections. We thank Michael Warner and Rohini Singh for comments that improved the manuscript. Footnotes†Co-first authors. Electronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.5007179. References
Page 10The open ocean is the largest habitable volume on earth and poses particular difficulties for the organisms living there. At mesopelagic depths (200 m–1000 m), the intensity of downward radiance is approximately 200 times that of upward radiance [1], so opaque organisms––even those with white ventral surfaces––create a silhouette that can be detected from below. To counteract this, a number of mesopelagic taxa (e.g. certain squid, crustaceans and fishes, including sharks) have arrays of ventral photophores that replace the downwelling light blocked by the body of the animal [2–6]. This form of camouflage, called counterillumination, is particularly common in stomiiform and myctophiform fishes, two of the most speciose and abundant orders of mesopelagic fishes [7]. The primary challenges for effective counterillumination are matching the intensity, spectrum and angular distribution of downwelling light. The ventral photophores of mesopelagic fishes are known to emit light with a spectrum that is a close match to the spectrum of downwelling light [6,8]. Additionally, at least two stomiiform species, Chauliodus sloani and Argyropelecus affinis, have photophores that use guanine reflectors to match the angular distribution of downwelling light [9]. The photophores of Stomiiformes and Myctophiformes are under neural control [10,11], and certain mesopelagic fishes (Dasyscopelus obtusirostris and Dasyscopelus spinosussensu Martin et al. [12], formerly Myctophum obtusirostre and Myctophum spinosum, respectively) adjust their ventral light emission in response to changes in downwelling light intensity over a range of up to 15 000-fold [6,13]. Despite evidence that counterillumination matches the intensity of downwelling light, little is known about the feedback system that mediates this process, given that these fishes are unable to see their own ventral photophores. It has been hypothesized that fishes may use a reference photophore, with an intensity correlated with the intensity of the ventral photophores, to emit light towards the eye (termed eye-facing photophore hereafter), allowing modulation of ventral bioluminescence until it matches the downwelling light viewed by the eye [4,14]. The potential use of such a reference photophore mechanism, however, has only been examined in one species, the myctophid Tarletonbeania crenularis [14]. To evaluate the role of an eye-facing photophore as part of a more general mechanism of regulating counterillumination, we characterized the presence/absence and orientation (eye-facing or not) of the orbital photophores of 36 species of stem group Stomiiformes in a phylogenetic context (we sampled 15 of 18 genera in the tree from Rabosky et al. [15]). In addition to those traits, we assessed the presence of an aphakic gap, a lensless section of pupil that allows for increased light capture, for the 14 species for which we were able to acquire photographs of fresh specimens. We then examined the morphology of the eye-facing photophores and surrounding tissues in one species in which the morphology of the retina is well characterized, the hatchetfish Argyropelecus aculeatus, to determine whether light from its eye-facing photophore reaches the eye, and to explore other potential functions of the emitted light. We focused on stem stomiiform fishes, members of the families Gonostomatidae (bristlemouths), Sternoptychidae (hatchetfishes) and Phosichthyidae (lightfishes). We chose these families because they are known to possess either a pre-, ant-, or suborbital photophore, but do not have the proliferation of cranial photophores seen in the crown group Stomiidae. The additional cranial photophores in stomiids suggests functional diversification, with many of these photophores hypothesized to function as searchlights, a process that does not necessarily require the same feedback as counterillumination [16,17]. Twenty-one specimens from 21 species were collected via high-speed rope trawl or 10 m2 MOCNESS in the Gulf of Mexico between 2009 and 2017. Immediately following collection, animals were placed in 10% formalin in buffered seawater and stored until use, at which point they were transferred to 70% ethanol. One specimen of A. aculeatus was collected via midwater Tucker trawl from Baltimore Canyon near 38.05° N 73.7° W and immediately placed in 2% glutaraldehyde in buffered seawater for use in histology (the A. aculeatus specimen collected previously in the Gulf of Mexico was stained for micro-CT). The remaining 15 specimens were acquired via loan from the Smithsonian Institution Museum of Natural History, Washington, DC, USA. One specimen per species was used because, while there are intraspecific differences in photophore patterning in some deep-sea fishes associated with changes in body size and population divergence [18–20], to our knowledge there are no shifts in the overall direction of photophores within a species (i.e. a photophore directed into the eye or away from the eye). Additional details on specimen collection, including museum acquisition numbers and trawl depth, can be found in electronic supplementary material, table S1. Of the 36 species considered here, 25 from 13 genera were included in a comprehensive molecular phylogeny of ray-finned fishes [15]. Those that were not found on the phylogeny were not included in the models of discrete trait evolution because they could not be unambiguously placed on the tree. The full tree was pruned to include only the species surveyed in this analysis using the packages ‘ape v5.3’ [21] and ‘phytools v0.6–60’ [22] in R [23]. To determine if there is a phylogenetic signal in the distribution of eye-facing photophores or ventral photophores that are capable of effective counterillumination, we calculated Fritz and Purvis's D statistic using 1000 permutations with the package ‘caper’ [24,25]. To consider the coevolution of ventral photophores capable of counterillumination and the eye-facing photophore, we fitted independent and dependent models of discrete trait evolution in BayesTraits v. 3.0.1 [26]. All parameters were set to their default values and the results of each model were compared using maximum likelihood. Given their proximity to the skin and the transparent tissue surrounding them, many eye-facing photophores could be identified under a stereo microscope (M5, Wild, Heerbrugg, Switzerland). For specimens (5/36) where the presence and orientation of an orbital photophore could not be inferred from standard imaging, it was characterized using micro-computed tomography (micro-CT). Prior to micro-CT scanning, whole fish were stained for two to five days in a 50 : 50 mixture of 2% aqueous solution of Lugol's iodine (J. Crows LLC, Ipswich, NH, USA) and the initial preservation medium (either 70% ethanol or 10% buffered formalin solution). Micro-CT was performed using a Nikon XTH 225 ST scanner at the Duke University Shared Materials Instrumentation Facility, which produced voxels with edge lengths between 1.2 µm and 4.1 µm with beam settings of 110 kV–190 kV and 43 µA–114 µA. We confirmed via dissection that micro-CT can be used to accurately identify both the presence of a photophore and its orientation because of differential staining of the lens and photocytes and the asymmetrical morphology of the photophore (figure 1). Figure 1. Micro-CT images of (a) Sternoptyx pseudobscura, (b) Argyropelecus hemigymnus and (c) Sternoptyx diaphana showing the presence and orientation of the photophores from iodine-stained specimens. The eye-facing photophore and the lens are shown in the inset panels. Eye-facing photophores and ventral photophores denoted by the arrows and brackets, respectively. Top: lateral view. Bottom: ventral view. All scale bars are 5 mm. (Online version in colour.)
A block of tissue from Argyropelecus aculeatus containing the photophore, part of the upper jaw, and the tissue between the cornea and the photophore was excised with a razor blade and partially dehydrated in 95% ethanol for 24 h. The tissue was embedded in glycol methacrylate plastic (GMA) to minimize tissue distortion (Technovit 7100, Kulzer GmbH, Hanau, Germany), and 2 µm thick sections were cut using a glass knife (Reichert-Jung, Leica, Wetzlar, Germany). Sections were stained with Picrosirius/Fast Green stain (0.3 g Sirius Red, 0.3 g Fast Green FCF, 300 ml saturated picric acid) for 3 h at 60°C to differentiate between tissue types. Following staining, samples were rinsed with deionized water, dried and mounted under a coverslip. Composite images were taken with a Zeiss Axiocam HRc digital camera on a Zeiss Axiophot microscope (Zeiss, Oberkochen, Germany). To further assess potential morphological specialization that may permit light from the eye-facing photophore to enter the eye, we assessed the presence and location of the aphakic gap for species in which we were able to acquire photographs of fresh specimens (14/36). We used fresh specimens only because many long-preserved specimens have distortions of the iris and tapetum that make assessment of the aphakic gap difficult. In total, we acquired aphakic gap data for 14 of the 36 species included in the study. Eye-facing photophores were present in all species with ventral photophores that are capable of counterillumination (34/36 species; for examples of light paths, see figure 2). Additionally, in at least A. aculeatus, S. diaphana and Maurolicus spp., the path of the light emitted from the eye-facing photophore intersects regions of high retinal cell density [27–29]. All eye-facing photophores were surrounded by a melanin layer on the side opposite the eye (i.e. anterior side for a photophore located in front of the eye) and the lateral face (figure 3; full trait data in electronic supplementary material, table S2). This pigment layer ostensibly absorbs light that is not propagating towards the eye and thus prevents the eye-facing photophore from being used as a searchlight, lure, or laterally directed signal. Conversely, we did not find eye-facing photophores in Cyclothone obscura and Sigmops bathyphilus, two primarily bathypelagic species [30]. While C. obscura has lost all photophores, S. bathyphilus maintains small masses of bioluminescent tissue known as secondary photophores and one short series of organized photophores on the ventral surface. These photophore arrays are unlikely to be sufficient for effective counterillumination, although it is possible they are used to disrupt the silhouette [31,32]. Figure 2. Light path from eye-facing photophore of three representative species. Micro-CT images from (a) A. hemigymnus, (b) S. diaphana and (c) M. weitzmani, showing zoomed-in side and top views of the eye-facing photophore and the eye. The dashed lines for all views are parallel to face of the photophore closest to the eye. Arrows are perpendicular to the dashed lines and denote the predicted light path. The top view for S. diaphana is in the plane of the dorsal surface of the photophore, not in the plane of the most-dorsal surface of the body, because the photophore is obscured dorsally by the musculature (note––the musculature does not extend into the light path between the photophore and the eye). (Online version in colour.)
Figure 3. (a) Phylogeny of sampled species of Gonostomatidae, Sternoptychidae and Phosichthyidae pruned from a comprehensive molecular phylogeny of fishes [15]. Colours indicate the presence or absence of the ventral photophores, the eye-facing photophore and an aphakic gap (where available). Only 25/36 species are included in the tree, but the remaining 11 species all exhibit ventral photophores capable of counterillumination and eye-facing photophores (Ariaophos eastropas, Agyripnus atlanticus, A. brocki, A. ephippiatus, Argyropelecus pacificus, Cyclothone braueri, Maurolicus muelleri, Polyipnus aquavitus, P. nuttingi, P. spinifer, Vinciguerria poweriae). Dissecting scope images show examples of the eye-facing photophore surrounded by melanin in (b) Cyclothone pallida, (c) Argyropelecus lychnus and (d) Maurolicus japonicus. The eye-facing photophores are denoted by the white arrows. (Online version in colour.)
The phylogenetic distributions of ventral photophores capable of counterillumination and the eye-facing photophore, when considered independently, were not significantly different than expected under random phylogenetic structure or Brownian motion (BM) (D = −0.431; Random: p = 0.164; BM: p = 0.635); however, this may be an underestimate of the phylogenetic signal because it does not include the 11 species that could not be unambiguously placed on the tree. When considering the coevolution of the two traits, a model of dependent evolution of eye-facing photophores and ventral photophores capable of counterillumination provided a better fit for the distribution of both traits than a model treating both traits independently (chi-square: p = 0.007). Our findings indicate that the ventral photophores and eye-facing photophores evolved in a correlated fashion, where the rate of transitioning to each presence–absence state for each trait depends on the presence–absence state of the other. This result is in line with the expected outcome given the one-to-one relationship between the presence of ventral photophores capable of counterillumination and eye-facing photophores (figure 3; electronic supplementary material, table S2). It should be noted, however, that there are only two transitions between character states, and no species that have one trait but not the other. We found that the eye-facing photophore in A. aculeatus is enclosed on the dorsal and ventral sides by two layers of connective tissue that flare away from the light-emitting end of the photophore to allow light propagation towards the eye (figure 4). Additionally, the lateral and anterior sides of the photophore are shielded by a layer of melanin that is several granules (1 µm–21 µm) thick. This pigment layer, like the layer found in other species with an eye-facing photophore, prevents the use of the orbital photophore as a searchlight for finding prey, a proposed function of the orbital photophores of many stomiid dragonfishes (the crown group of the order) [33]. Between the eye-facing photophore and the cornea of the eye is a region of tissue that is transparent in fresh specimens, translucent in preserved specimens and does not stain with Fast Green, suggesting little protein content (and thus presumably minimal scattering). Light exiting the eye-facing photophore and travelling through this transparent tissue would be absorbed by the iris in the tubular eyes of other deep-sea fishes (e.g. Opisthoproctussoleatus) but is able to pass through a dip in the nasal edge of the iris that is reported by Collin et al. [27]. It then presumably passes through the lens and falls on the accessory retina, likely as an unfocused spot because the distance between the accessory retina and the lens is far less than 2.55 times the lens radius (the predicted focal length for fishes from Matthiessen's ratio [34]; figure 4). Figure 4. Digital illustration of Argyropelecus aculeatus showing the morphology of the eye and eye-facing photophore constructed from micro-CT, histological measurements, and previously described retinal morphology [27]. Box 1: the eye-facing photophore (p) emits light that passes through a volume of transparent tissue (tt) and through a dip in the iris (i) reported by Collin et al. [27]. After travelling through the lens, light strikes the accessory retina (ar) while downwelling light illuminates the main retina (r). Box 2: histological section through the middle of the eye-facing photophore shows photocytes (p), the photophore lens (l) and a layer of melanin (m) surrounding three sides of the photophore, preventing light from escaping anteriorly or laterally. Note, there is no reflector or filter in the eye-facing photophore. Additionally, collagen layers (c) flare out from the end of the photophore that the light exits. Once leaving the photophore, light travels through tissue (tt) that is largely transparent and does not stain with Fast Green. Scale bars: box 1, 500 µm; box 2, 100 µm. (Online version in colour.)
We found aphakic gaps in eight of the 14 species for which we could get photographs of fresh specimens (trait data: figure 3). In A. aculeatus, A. affinis, A. sladeni, I. ovatus and V. tripunctulatus the aphakic gap consisted of a dip in the nasal edge of the iris that exposed the lens to the eye-facing photophore. Typically, without this dip, the pigmented iris found in other fishes with tubular eyes would block light coming from the photophore. In S. diaphana and S. pseudobscura, the aphakic gap is located ventral to the lens. The eye-facing photophore in these species is located on the orbit posterior and dorsal to the lens and directs light ventrally and anteriorly (figures 1 and 2). The ventral aphakic gap in Sternoptyx spp. aligns with the light path from the photophore, allowing the emitted light to enter the eye. Additionally, at least in S. diaphana, the ventral part of the retina where the photophore light would fall has a high density of cells in the ganglion cell layer [28]. In P. clarus, the aphakic gap is located along the ventral margin of the lens and the eye-facing photophore is directed posteriorly and ventrally, roughly aligning with the location of the aphakic gap. The remaining six species (C. braueri, C. obscura, C. pallida, P. mauli, S. elongatus and V. poweriae) do not possess an aphakic gap. There were no species for which we recorded an aphakic gap that did not align with the direction of the eye-facing photophore. Here, we show that 34 of 36 sampled species of stem Stomiiformes (of the 53 found on a phylogeny) [15] are counterilluminators based on the presence of ventral photophores, and that these species have a photophore directed into the eye that is pigmented in a way that rules out a searchlight function or use as a laterally directed signal. While behavioural experiments directly linking the eye-facing photophore to counterillumination regulation were not possible because mesopelagic fishes caught in trawl nets are typically deceased or moribund, we found that the distribution of ventral photophores and the eye-facing photophore are best explained by a dependent model of discrete trait evolution. That is, considering complete rows of ventral photophores (a proxy for counterillumination) and the eye-facing photophore as correlated traits provides a higher maximum likelihood than treating the two traits independently. Additionally, a number of species possess morphological specializations such as aphakic gaps that permit light from the eye-facing photophore to enter the eye and strike the retina. Further supporting the role of the eye-facing photophore in regulating counterillumination are reports of similar photophores in distantly related counterilluminating species. Lawry [14] found that the myctophid Tarletonbeania crenularis has a small photophore dorsal to the eye that casts light on the ventral part of the retina. Also, barracudinas (Paralepididae) in the genus Lestrolepis have a small organ anterior to the eye that is pigmented on the anterior and lateral sides that is presumed to be bioluminescent [35]. Finally, etmopterid sharks have small photophores dorsal to the eye that may cast light on the retina in a similar manner to T. crenularis [36]. While future work is necessary to systematically describe the distribution and morphology of eye-facing photophores in these groups of counterilluminating fishes, the widespread distribution of eye-facing photophores suggests that this mechanism of counterillumination regulation could extend beyond Stomiiformes. The only suggested alternative hypothesis for the function of the eye-facing photophore is that, by shining light in the eye, it changes the adaptation state to ‘prime’ the eye (i.e. lower its sensitivity) and allow the fish to see any potential prey that are illuminated by a searchlight photophore without blinding itself. Priming the eye is unlikely to be the function of the eye-facing photophore in these families because they do not possess a searchlight photophore. Given this, there is no likely alternative function, to our knowledge, for a photophore that reduces sensitivity and increases glare, especially in a light-limited environment characterized by ocular adaptations that increase sensitivity (e.g. large pupil, tubular shape, long photoreceptors, spatial summation, aphakic gaps and accessory retinae [28]). Further investigation may reveal associations between these other ocular adaptations and eye-facing photophores. In sum, our results suggest that stem stomiiformes use a photophore directed into the eye to calibrate and thereby aid in the regulation of the intensity of their ventral bioluminescence for counterillumination camouflage. All data are available in the text and the electronic supplementary material. A.L.D and S.J. conceived the work; T.S. provided specimens; S.J. acquired funding; A.L.D. and W.M.K. did the histology; A.L.D performed the dissection, micro-CT, data analysis and wrote the manuscript. All authors were involved in the discussion of the results and revision of the manuscript. Authors declare no competing interests. Funding for this work was provided by the Duke University Department of Biology, the National Science Foundation (IOS 1557754 to W.M.K.) and was made possible in part by a grant from The Gulf of Mexico Research Initiative. Data are publically available through the Gulf of Mexico Research Initiative Information & Data Cooperative (GRIIDC) at https://data.gulfresearchinitiative.org (doi:10.7266/N7VX0DK2; doi:10.7266/N7R49NTN). A.L.D. conducted this research with Government support under and awarded by DoD, Army Research Office, National Defense Science and Engineering Graduate Fellowship (NDSEG). We thank Justin Gladman and Dr. Carlos Taboada for assistance with micro-CT, and Jesse Granger, Sarah Solie, Dr Eleanor Caves and Dr Robert Fitak for helpful comments on an earlier version of the manuscript. We also thank the crew of the R/V Hugh R. Sharp, R/V Point Sur and M/V Meg Skansi for assistance at sea. FootnotesElectronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.4994588. References
Page 11Humanity’s unparalleled cultural and technological sophistication has been widely attributed to our ability to not just share information, but continually build upon it as well [1,2]. This process, called cumulative cultural evolution (CCE), has resulted in knowledge and technology that no single generation could produce on its own. However, despite extensive evidence of culture in a wide range of species [3], non-human animals have demonstrated only a limited capacity for CCE. Not only has observational evidence proved scarce and contentious [4], but experiments have shown that CCE can be surprisingly difficult to evoke even in closely related primates [5,6]. While some examples have been elicited in various species [7], these often involve extensive human intervention and remain comparatively modest. This raises a question that has perplexed biologists, psychologists and anthropologists alike: what makes humans, if not unique in our capacity for CCE, uniquely adept at producing it? CCE arises when social learning preserves information between generations, allowing individual learning or lucky errors in transmission to refine it [8]. This process probably depends far more heavily on how reliably information is preserved than on how efficiently it is refined, because the more knowledge accumulates, the more there is to rediscover or reinvent when transmission fails. Theoretical models explicitly support this idea [9] and often find that transmission fidelity must pass a threshold for culture to accumulate [10] (though see cultural attractor theory [11] for an alternative view). Notably, humans transmit information with exceptionally high fidelity by not only communicating through language, but also imitating more accurately [6] and robustly [12], leveraging a more sophisticated theory of mind [13], showing natural inclinations towards pedagogy [14] and practising a far wider range of teaching behaviours [15]. This has led to the view that CCE relies on accurate but specialized forms of social learning at which humans are particularly adept [2,16,17]. Precisely what social learning mechanisms underlie CCE remains unclear, however. Researchers have long emphasized the role of imitation (process-copying) and teaching, drawing sharp contrasts with less accurate forms of social learning like emulation (product-copying) [2,16,17]. On this front, transmission chain and laboratory microsociety studies have yielded contradictory results. Some have found that imitation and emulation both support CCE [18–20], while others suggest that emulation is insufficient [21,22]. To complicate matters further, studies emphasizing ecological validity have found that even imitation fails to preserve early stone tool manufacturing (knapping) techniques. Teaching through gesture [23] or even language [24] may thus be critical to human-like CCE. Given this empirical ambiguity, it may be useful to draw a functional distinction between high-fidelity social learning that supports CCE and low-fidelity social learning that does not, regardless of what the underlying mechanisms turn out to be. Bayesian models drawn from work on language evolution have shown how this can be achieved [25,26]. These reveal that when social learning is captured as sampling and inference, it is too low-fidelity for knowledge to accumulate [25]. However, when social learning is captured as the direct transmission of beliefs [25] or information about those beliefs [26], it can give rise to CCE. A Bayesian framework thus delineates between these two types of learning in a mechanism-agnostic way. Thus far, such models have largely been used to study cultural evolution in transmission chains. However, they also present an opportunity to address a more fundamental question: why would biological evolution produce high-fidelity social learning in some species and not others? Early models showed that CCE cannot explain the evolution of accurate transmission, because CCE would take many generations to pay for this upfront investment [16]. As a result, much of the CCE literature has taken such transmission for granted and focused on other factors instead, such as demography, social connectedness, transmission biases and filtering of maladaptive traits [7]. Here, we develop a Bayesian model that contrasts the evolution of high- and low-fidelity social learning directly. Doing so reveals that high-fidelity transmission evolves under different conditions than social learning that spreads culture but does not refine it. Consider a population facing an adaptive problem that involves estimating a set of parameters, Θ = {θ1, θ2, …, θx}, where each θ takes some value between 0 and 1. Beliefs about each θ are encoded as a probability distribution, p(θ), that describes which values an individual deems likely and which it does not. For example, if Θ encapsulates knowledge about constructing a spear, then elements θ1, θ2 and θ3 could represent the spear’s ideal length, diameter and centre of gravity (where each characteristic is normalized to fall between some minimum plausible value, represented by θ = 0, and some maximum plausible value, represented by θ = 1). Similarly, Θ could encode knowledge about knapping, where θ1 through θx represent the ideal striking platform angle, flaking surface concavity, distance from the edge, amount of force to apply, etc. Alternatively, Θ could capture how much time and effort to devote to one food patch (θ1) as opposed to another (θ2) and thus encode a foraging strategy. Learning occurs when beliefs, p(θ), change in response to new data, d, resulting in an updated set of beliefs, p(θ|d). This is modelled as Bayesian inference, p(θ|d)=P(d|θ)p(θ)∫01P(d|θ)p(θ)dθ,2.1 where posterior beliefs, p(θ|d), are a product of prior beliefs, p(θ), and the likelihood of observing the data if those priors are true, P(d|θ). Bayesian inference thus takes a learner’s beliefs and updates them with new data, such that surprising data change beliefs to a greater extent. The denominator is simply a normalizing term, which ensures that probabilities integrate to 1.Beliefs about each θ follow a beta distribution and data, d, consist of either n samples drawn from the environment or m samples drawn from the population. After learning, individuals select the most plausible value of θ as their estimate. This is the posterior distribution’s mode, θ^MAP=argmaxθp(θ|d),2.2 which can be calculated directly from a beta distribution’s shape parameters: θ^=(α−1)/(α+β−2). This makes our model analytically tractable, because it allows us to reason in terms of the data individuals observe rather than the resulting distributions.Taken together, these estimates shape the individual’s trait. This trait’s efficiency is defined by z=x−∑i=1x|θi−θ^i|,2.3 where Θ^={ θ^1,θ^2,…,θ^x} are the individual’s estimates after learning and x is the trait’s complexity (the set’s cardinality). When estimates lie close to their ideal values, absolute error is minimized and trait efficiency approaches z = x. Conversely, when estimates lie far from their ideal values, error is maximized and trait efficiency is low (z = 0 in the extreme case where each θ and θ^ take opposite values of 0 and 1).This formulation makes several simplifying assumptions. First, a trait’s maximum efficiency grows linearly with trait complexity (x). We will see later that this assumption can be weakened to include other growth rates (e.g. logarithmic), subject to some constraints. Second, each trait has a single optimal variant (a unimodal adaptive landscape), which is not necessarily true in complex domains like tools [27]. Third, each parameter is independent, with the ideal value of one θ having no effect on the ideal value of others. In reality, such contingencies do occur, for example, in knapping [28]. In our model, priors reflect common intuitions about θ, whose influence diminishes with learning. These may arise through similarities in genes, ontogeny, previous experience, etc. For example, if individuals share only weak intuitions about the ideal length of a spear, some novices could make long spears while others make short ones. Alternatively, if individuals share strong biases about the amount of force to apply when knapping, novices could consistently overestimate this parameter. In fact, such a pattern has been observed in experiments [28]. We use an asterisk to denote prior estimates, θ^∗, and trait efficiency, z*. An adaptive problem’s difficulty can be defined as the average distance between a parameter’s ideal value and the prior estimate, f=1x∑i=1x|θi−θ^i∗|. When problems are hard, the optimal trait is unintuitive and a lot of learning is needed. Conversely, when problems are easy, efficient solutions are obvious, and there is little or nothing to learn. This could be due to luck, shared relevant experience or even because evolution has yielded an innate adaptive behaviour [29]. Individual learning involves interacting directly with the environment, through observation, exploration or trial-and-error. We formalize this as sampling a random variable X, where E[X] = θ. For example, in foraging, a sample could indicate whether a given food patch was productive or unproductive, such that X ∼ Bernoulli(θ). Alternatively, in knapping, a sample could indicate the distance from the platform edge that produced a viable flake. Distances closer to the ideal could be more likely to succeed, such that X∼N(θ,σ2). Let n be the average number of samples per parameter. The average individual learner’s estimate is thus θ^¯I=θn+ θ^∗vn+v,2.4 which reflects the combined influence of the environment (θ) and the prior ( θ^∗). Note that the relative weight placed on the prior, v > 0, can be understood as the number of ‘virtual samples’ that would be needed to form that distribution. Because more genuine samples are needed to overcome stronger priors, v serves as a measure of conservatism.Each sample comes at some cost, c ≥ 0, which represents time, energy, opportunity cost, risk of injury or predation, etc. More sampling yields a more efficient trait, but comes at a greater overall cost, cnx. For example, making three spears gives more insight into the ideal length of a spear than making two would, but requires additional time, effort, material and risk. The average individual learner’s fitness is thus ω¯I=ω0+z¯I−cnx,2.5 where ω0 represents aspects of fitness unrelated to learning.In Bayesian inference, each sample improves accuracy less than the preceding one. Because the per-sample cost (c) is invariant, this captures the notion of diminishing returns. The optimal learning rate, which maximizes expected utility and fitness, is n= fvc−v.2.6 Intuitively, individuals learn more when doing so is inexpensive (low c) and problems are difficult (high f). Conservatism (v) has a more complicated effect. When individuals are highly conservative, it is not worth collecting many samples, because beliefs barely change with new data. Likewise, when priors are extremely diffuse, few samples are needed to sway the learner. Sampling peaks when behaviour is flexible and priors are weak, but not so weak that individuals show no skepticism towards surprising data.Combining equations (2.3), (2.4) and (2.6) gives the average individual learner’s trait efficiency z¯I=x(1−cfv).2.7 Because v > 0, individual learning cannot reliably acquire the optimal trait, z¯I=x, unless learning is free (c = 0) or the initial trait is already optimal (f = 0). If learning is costly and difficult, then individual learning only partially improves the trait and CCE is needed to reliably acquire the ideal variant.In low-fidelity social learning, individuals learn about the environment by observing others’ behavioural outcomes. For example, seeing many long spears but few short ones is indirect evidence that longer spears are more effective. In reality, behavioural outcomes often fail to accurately reflect beliefs, resulting in incomplete information and errors in inference [30]. To capture this notion, learners do not sample an estimate directly, but rather a random variable Y, where E[Y]=θ^. For instance, if a demonstrator tries to build spears of length θ^, errors in production may result in some shorter and some longer ones, such that Y∼N(θ^,σ2). Let m be the average number of samples per parameter. The average low-fidelity social learner’s estimate is thus θ^¯L=θ^¯m+θ^∗vm+v,2.8 which reflects the combined influence of social information (θ^¯) and the prior (θ^∗). We confirm in electronic supplementary material, §1.1 that such social learning does not support CCE, because it cannot improve average trait efficiency over time when combined with individual learning.Each sample comes at some cost, k ≥ 0, which represents the expenditure and risk involved in surveilling others. Collecting additional samples allows learners to more faithfully reproduce the average trait, but comes at a higher overall cost, kmx. The average low-fidelity social learner’s fitness is thus ω¯L=ω0+z¯L−kmx.2.9 As in individual learning, sampling yields diminishing returns. The optimal social learning rate is m=vkx∑i=1x|θ^¯i−θ^i∗|−v,2.10 though such learning should be avoided entirely, m = 0, if others haven’t improved on the initial trait, z¯≤z∗. More effort is devoted to learning when doing so is inexpensive (low k) and there is more knowledge to acquire (the summed term is large).Combining equations (2.3), (2.8) and (2.10) gives the average low-fidelity social learner’s trait efficiency, z¯L=z¯−kvx∑i=1x|θ^¯i−θ^i∗|.2.11 Such learning cannot reliably preserve others’ knowledge, z¯L=z¯, unless learning is free (k = 0) or there is nothing to learn (the summed term is 0). Otherwise, some knowledge is lost in transmission and supplanted by prior beliefs [25].High-fidelity social learning involves faithfully reproducing an existing trait, which we formalize as copying another individual’s estimates. One way this could happen is if a learner adopts identical underlying beliefs [25]. For example, language or gesture could convey everything a teacher knows about where to aim blows when knapping. Alternatively, a learner could adopt beliefs that are merely compatible with the observed trait (i.e. different distributions with the same posterior mode). For instance, accurately imitating a demonstrator’s construction process could yield spears of the same average length, but subtly different beliefs about the relative efficiency of shorter or longer ones. In either case, the average high-fidelity social learner’s estimate is identical to that of the population, θ^¯H=θ^¯, as is its trait efficiency, z¯H=z¯. We confirm in electronic supplementary material, §1.2 that such social learning supports CCE, because it can improve average trait efficiency over time when combined with individual learning. Each parameter individuals copy comes at some cost. Thus far, we have assumed that social and individual learning rely on the same cognitive mechanisms [31] and that the evolution of social learning primarily reflects changes in attention and motivation. However, high-fidelity transmission may involve more specialized and cognitively demanding forms of social learning [2,16,17]. For example, if it involves accurate imitation, then it may require specialized neural machinery for parsing and reproducing bodily actions that has undergone significant elaboration in the hominin lineage [12,32]. Alternatively, if it involves human-like teaching, then it may require the capacity for gesture or even language [33]. Though some researchers argue that high-fidelity transmission is as much a product of cultural as of biological evolution [34], some genetic endowment is clearly needed, even if this consists of a mere ‘start-up kit’ that is later refined through culture [35]. That being said, the addition of brain tissue is notoriously energetically expensive, particularly during development [36]. The cost of high-fidelity social learning may thus consist of two components: a dynamic component, gd, that reflects the expenditure and risk involved in employing such learning; and a static component, gs, that reflects the cost of developing and maintaining it. This gives an overall cost gdx + gs ≥ 0, where the dynamic cost grows with how extensively learning is employed (x), but the static cost is invariant. To capture both components as a single per-parameter cost, we define g = gd + gs/x. The average high-fidelity social learner’s fitness is thus ω¯H=ω0+z¯−gx.2.12 To contrast the evolution of high- and low-fidelity social learning, we track the fate of rare social learning mutants in a monomorphic population of individual learners, where z¯=z¯I and ω¯=ω¯I. Social learning goes extinct if these mutants’ average fitness (ω¯L or ω¯H) falls below that of the resident type (ω¯I). Conversely, social learning evolves if these mutants have higher fitness and their invasion results in either fixation or coexistence (a dimorphic equilibrium). We do not consider dimorphic resident populations, because for our purposes the effects would be fairly straightforward. Namely, the resident population’s average trait efficiency (z¯) would decrease as low-fidelity social learning became more common, which is equivalent to a monomorphic case where high-fidelity transmission is more costly (i.e. g is Δz¯/x higher). While social learning is often subject to frequency-dependent selection [37], this does not concern us for two reasons. First, high-fidelity social learning’s fitness is not frequency-dependent at all, because it simply maintains the population’s average trait (z¯H=z¯) and this trait’s efficiency does not change over time (cf. [37]). Any fitness advantage it has as a rare mutant thus persists until fixation. Second, while low-fidelity social learning’s fitness is frequency-dependent, this never brings about its extinction. Rather, as such mutants become more common, their average trait efficiency declines until their fitness equalizes with that of the resident type, ω¯L=ω¯I, resulting in a dimorphic equilibrium. We show in electronic supplementary material, §2 that such an equilibrium exists and is stable whenever they invade. Social learning cost is central to our analysis, because it reveals both when a given type of learning could conceivably pay for itself and when it is best equipped to do so. Setting ω¯L=ω¯I gives the maximum per-sample cost of low-fidelity social learning, kmax=(cv− f)(2 f−cfv+cv−2 f)v,3.1 and setting ω¯H=ω¯I gives the maximum per-parameter cost of high-fidelity social learning,gmax=cfv−cv.3.2 At or above these values, such learning no longer confers a fitness advantage. Identifying when kmax > 0 or gmax > 0 thus reveals the minimum requirements for social learning to evolve. More importantly, conditions that maximize kmax or gmax reveal when such learning withstands the broadest possible range of costs and is thus most likely to evolve (though such conditions do not necessarily maximize its prevalence in the population, learning rate, etc.).For social learning to evolve, it must either improve on the average trait or reduce the cost of acquiring it [38]. Although transmission errors can yield a superior trait, lucky mistakes are no more likely to be observed than unlucky ones (cf. [30]). Therefore, social learning must reduce cost. Setting ω¯L>ω¯I reveals that low-fidelity social learning evolves when its savings in cost exceed its average loss in trait efficiency cnx−kmx>z¯I−z¯L.3.3 The more errors in transmission, the larger the necessary savings. By contrast, high-fidelity social learning makes virtually no errors in transmission. It thus evolves (ω¯H>ω¯I) when it offers nearly any savings in costcnx−gx>0.3.4 (In reality, there will probably always be some slight, non-zero level of error to overcome.) Taken together, equations (3.3) and (3.4) imply that high-fidelity social learning tolerates a greater overall cost by maintaining more efficient traits.Trait complexity can be eliminated from equation (3.3), because each term grows linearly with x. Doing so yields the equivalent expression cn−km>| θ^¯I−θ^¯L|, which implies that low-fidelity social learning is as likely to evolve when traits are simple as when they are complex. The same is not true of equation (3.4), once we break cost g down into its static and dynamic components. Instead, the evolution of high-fidelity social learning requires crossing a threshold in trait complexity, x > gs/(c n − gd), which increases with the cost of both having (gs) and employing (gd) such learning. Note that this result is not contingent on our assumption that trait efficiency and learning cost grow linearly with respect to trait complexity, but rather that they grow at the same rate. For example, x could still be eliminated from equation (3.3) if efficiency and cost both grew logarithmically (e.g. if increased complexity yielded diminishing returns in efficiency, but learning one parameter made it easier to learn others). Social learning can only evolve (kmax > 0 or gmax > 0) when there is knowledge to acquire, n > 0. However, different types of social learning benefit from vastly different individual learning rates (figure 1a). This can be seen by finding the values of n that maximize kmax and gmax (after first simplifying these expressions by using equation (2.6) to substitute c = f v/(n + v)2). Doing so reveals that low-fidelity social learning is most likely to evolve when the individual learning rate is low, n = v/3, and beliefs are driven mostly by prior expectations. By contrast, high-fidelity transmission is most likely to evolve when the individual learning rate is much higher, n = v. Figure 1. Resilience of high- and low-fidelity social learning as a function of: (a) the individual learning rate, (b) individual learning cost, (c) problem difficulty and (d) conservatism. Dotted lines indicate when social learning is most likely to evolve. High-fidelity transmission evolves when individual learning is comparatively (a) plentiful and (b) inexpensive. Its evolution may also depend on confronting particularly (c) challenging adaptive problems, because it accrues the benefits of increased problem difficulty more slowly. Finally, while all social learning benefits from (d) behavioural flexibility, high-fidelity transmission could benefit from higher levels of conservatism if these stimulate rather than depress individual learning. Parameters: c = 0.005, f = 0.5, v = 12.
Social learning cannot evolve when individual learning is free, c = 0, because it confers no savings. Similarly, it cannot evolve when individual learning is too expensive to engage in, c ≥ f/v, because there is nothing to learn. Between these two extremes, however, different individual learning costs favour different types of social learning (figure 1b). Low-fidelity transmission is most likely to evolve when individual learning is relatively expensive, c = 9f/(16v), whereas high-fidelity transmission benefits from much cheaper individual learning, c = f/(4v). In fact, the latter regime represents a 5/9≈56% reduction in cost. Adaptive problems must be sufficiently difficult, f > cv, for learning to evolve. Below this threshold, learning is not cost-effective, because the optimal trait is highly intuitive. Harder problems favour social learning in particular, which becomes more resilient as difficulty increases: ∂kmax/∂f > 0 and ∂gmax/∂f > 0. Social learning is thus most likely to evolve when problems are as difficult as possible, f = 1. That being said, high- and low-fidelity transmission react differently to increases in difficulty (figure 1c). Normalizing kmax and gmax by their maximum values reveals that kmax/kmax|f=1 > gmax/gmax|f=1 over c v < f < 1. In other words, low-fidelity transmission accrues the benefits of increased problem difficulty sooner. Larger increases are thus needed for high-fidelity transmission to reap comparable rewards (i.e. a proportional increase in resilience against cost). Learning can only evolve when the level of conservatism falls below v < f/c. Stronger priors make learning uneconomical, because updating beliefs involves collecting too much data. Low-fidelity transmission always benefits from reduced conservatism, ∂kmax/∂v < 0, and is thus most likely to evolve when priors are as diffuse as possible (low v). Although high-fidelity transmission also benefits from behavioural flexibility (figure 1d), its ideal level of conservatism is somewhat higher, v = f/(4c). This value is ideal because it maximizes the individual learning rate. A longstanding question about CCE is why humans acquired this capacity, which appears diminished or absent in other species. Given the importance of transmission fidelity [9], one explanation is that CCE relies on powerful but specialized forms of social learning at which humans are uniquely adept [2,16,17]. By characterizing social learning in terms of its ability to support CCE rather than specific underlying mechanisms, we find that high-fidelity transmission evolves under different conditions than less accurate social learning. Specifically, high-fidelity transmission is most likely to evolve when: (1) social and (2) individual learning are inexpensive, (3) traits are complex, (4) individual learning rates and (5) problem difficulty are high, and (6) behaviour is flexible. Low-fidelity transmission differs in many respects. Not only is it most likely to evolve when individual learning is (2) costly and (4) infrequent, but it is also more robust when (3) traits are simple and (5) problems are easy. If conditions favouring the evolution of high-fidelity transmission are stricter (3 and 5) or harder to meet (2 and 4), this could explain why social learning is common across species, but CCE is rare. Comparative analyses suggest that reliance on social learning covaries with brain size in primates [39,40]. Because the hominin brain has undergone several large evolutionary expansions [41], high-fidelity social learning may require the addition of costly brain tissue [16]. Our model suggests that one way to compensate for this increased expenditure would be to lower other costs associated with social learning. This could be achieved in several ways. First, social tolerance and grouping could provide easier, safer and more frequent opportunities to learn from others. In support of this view, sociability has been found to covary with reliance on social learning both within humans [42] and across primates [40]. Second, extended juvenile periods could free up time for social learning [43] without forgoing the opportunities in reproduction and resource acquisition available to an adult. Third, proactive prosociality could promote teaching [44]. Teaching, in this case, does not necessarily refer to the varied and cognitively complex forms it takes in humans [15,33], but rather to any instance where individuals modify their behaviour to foster others’ learning [44]. Pedagogy could thus drive its own evolution, with more elaborate forms of teaching evolving in response to this reduction in cost. Another way to offset the added cost of high-fidelity transmission would be through higher intake [36]. In line with previous models, we find that accurate social learning tolerates a greater overall cost precisely because it yields more efficient traits [16]. We build on this insight by allowing trait efficiency to grow with trait complexity. Though this relationship is not universal (e.g. simplifying a trait could make it more efficient), complexity is often indicative of improvement. For example, as knapping techniques became more elaborate and hierarchically structured, this resulted in better tools [45]. Following this assumption, we find that high-fidelity social learning is more likely to evolve when traits are complex, because the payoffs in trait efficiency dwarf the cost of developing and maintaining such learning. Unlike other species, early hominins may have crossed a threshold in trait complexity that allowed accurate transmission to evolve. This initial complexity may have arisen for reasons other than social learning, for example because encephalization allowed for more sophisticated action sequences [3]. This explanation is consistent with the archaeological record. Stout & Hecht [32] note that the first stone tools (3.3 Ma) saw only intermittent use and that even the early Oldowan technocomplex (2.6–2.0 Ma) gives the impression of being at the limits of hominin ability. Though the existence of local traditions suggests that Oldowan techniques were culturally transmitted [45], there is a conspicuous lack of evidence for CCE until much later on [46], following significant increases in brain size [41]. During this early period (and perhaps considerably beyond it [47]), social learning seems to have spread and maintained but not significantly refined the manufacture of tools. Not only is there no clear evidence of high-fidelity transmission [46] but the observed cultural dynamics closely align with those found in our model when individual and low-fidelity social learning are combined (electronic supplementary material, §1.1). Namely, a steady state emerges where average trait efficiency remains stable, but knowledge is repeatedly lost and rediscovered (socially mediated serial reinnovation [48]). In short, rather than high-fidelity social learning spreading and maintaining early lithic technologies, their relative complexity may have instead facilitated its evolution. The putatively high cost of accurate transmission is only one of the potential impediments to its evolution. Theory suggests that low individual learning rates could also play a role [38]. In line with this view, we find that much higher rates may be needed for the evolution of high- rather than low-fidelity social learning. Notably, the hominin lineage is characterized by large brains and high general intelligence, both of which are predictive of innovation rates in primates [40]. If few species are sufficiently prolific individual learners, this could explain why accurate transmission is rare. Of course, this raises the question of how adequate individual learning rates could be achieved in the first place. The most obvious way to stimulate individual learning is to reduce its cost. Previous theory [1] and experiments [42] warn that doing so can undercut social learning, however. While we find support for this view, we also find that high-fidelity transmission nevertheless benefits from such reductions. In practice, many of the same factors that mitigate the cost of social learning could do so for individual learning as well. First, grouping could reduce the cost of exploration by allowing individuals to diffuse the associated risks [49]. Second, extended juvenile periods could offer more time for not just social learning, but individual learning as well [43]. Costs borne by juveniles in protected environments, where others provide food, shelter and predator detection [43], would be especially affected. Finally, even teaching could play a role in the form of opportunity scaffolding, where a teacher does not necessarily demonstrate a behaviour, but rather furnishes students with easy and safe opportunities to learn on their own [50]. Another way to promote individual learning is by facing more challenging adaptive problems. We find that the evolution of high-fidelity social learning may involve confronting particularly difficult social, ecological and technological challenges (i.e. problems where optimal traits fall far outside the ‘zone of latent solutions’ [17]). There are several reasons to think that hominins confronted such problems. First, because bipedalism allows hominins to cover far larger geographical ranges than other primates, with lifetime home ranges several orders of magnitude greater than those of chimpanzees [51], individuals were likely subjected to greater variability in environmental conditions, available resources, potential threats, etc. If behaviour that is adaptive in one setting is non-adaptive in others, then problems may more frequently require unintuitive solutions. Second, an exceptionally large proportion of the hominin diet consists of high-quality foods [52], such as those procured through hunting, extractive foraging and confrontational scavenging [36]. Compared with foods consumed more regularly by other primates, these are skill intensive and difficult to obtain [52]. Finally, new ways of thinking, interacting with others and leveraging technology undoubtedly presented novel problems of their own. This probably resulted in a unique and challenging cognitive, cultural and technological niche, which further shaped the course of our evolution [32]. Lastly, it is worth commenting on the role of conservatism. A striking empirical finding is that chimpanzees suffer from remarkable functional fixedness and behavioural conservatism, which are thought to contribute to the paucity of CCE in this species [5,6]. We find that conservatism impedes CCE insofar as it disfavours investment into social learning. However, we also find that high-fidelity social learning could benefit from higher levels of conservatism if these stimulate rather than depress individual learning. For conservatism to impede the evolution of accurate transmission in particular, some additional assumption must be invoked, namely that such transmission also happens to be comparatively expensive. Individually, our criteria for evolving improved fidelity of transmission seem simple: mitigating the cost of learning, confronting harder adaptive problems, acquiring more complex traits, etc. However, our model emphasizes that meeting any one of these criteria is not necessarily sufficient. For example, even if migration exposes individuals to less intuitive problems, learning could still be too expensive. Similarly, even if grouping lowers the cost of learning, traits could still be too simple. In short, humans probably evolved high-fidelity social learning not by meeting any one (or more) of these criteria perfectly, but by meeting all of them well enough. This article has no additional data. M.M. and T.R.S conceived the project. M.M. developed the model, performed the analysis and wrote the manuscript. T.R.S provided manuscript revisions. We declare we have no competing interests. This work was supported by grants to M.M. (CGS-M and CGS-D) from the Natural Sciences and Engineering Research Council of Canada. FootnotesElectronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.4994594. References
Page 12It has been suggested recently that animal weapon exaggeration (the evolution of extremely enlarged or elaborate forms, beyond that which is usual among similar species) is more likely to occur in species where combatants face each other in duels [1]. This is because in duels, the outcome is expected to be more deterministic, with stronger or better-armed contestants consistently winning fights [1]. Thus, it is suggested that arms races are more likely to occur in duel-based systems, leading to the evolution of elaborate weaponry such as the exaggerated structures wielded by many male animals in fights over females (e.g. [2–7]). A key observation supporting this hypothesis is that among dung beetles (Scarabaeidae), those species that fight in duels tend to bear large and elaborate horns, while those that fight in skirmishes do not [8,9]. By contrast, systems where combatants face each other en masse (e.g. horseshoe crabs [10] and Dawson's burrowing bees [11]) may have less predictable outcomes, and weaker or less well-armed individuals may sometimes be victorious [1,12,13]. In such a scenario, selection may instead favour adaptations relating to energy efficiency, agility, endurance, learning or behaviour, instead of direct fighting power and extreme weaponry [14–17]. Duels may be just as important for weapon escalation in military technologies, for precisely the same reasons as described above in animal systems [1,18–20]. New technologies (e.g. battering rams on oared galleys, closeable gun ports on sailing galleons, machine guns on early aircraft) that aligned military vehicles in close-range one-on-one engagements (i.e. duels) have long been considered catalysts of military arms races [18,19]. If true, then this would point to an exciting parallel between animal and military forms of conflict. Experimentally testing for the importance of duels in human conflicts has obvious practical and ethical drawbacks, but there are additional non-animal systems in which there is severe conflict, without those obstacles. One example is computer-simulated warfare, where programmed combatants attempt to destroy each other in a digital behavioural medium. Helpfully, the realm of computerized war gaming has provided many highly tuned, adaptable and diverse conflict simulators with massive effort spent perfectly tuning the models for high-stakes E-sports competition [21–24]. In addition, the user interfaces of these programs are developed to allow detailed customization of factors such as arena layout, combatant characteristics, motivation, victory conditions and countless more. These advantages have made war games a favourite basis for research in artificial intelligence (AI), statistics, human cognition, machine learning and strategy (e.g. [25–29]). We feel that war games can have as much or an even greater benefit to understanding biological concepts such as animal contests and weapon evolution. Here, we set the programmed AI combatants provided by the real-time strategy (RTS) war game Starcraft 2 against one-another to test whether duels favour superior combatants disproportionately more than skirmishes. Similarly to the animals used extensively for contest research, including insects, spiders and crustaceans (e.g. [30–32]), the AI's senses are limited to their local environment, they operate according to a set of simple rules, they use resources from their environment in order to grow, and they can trade off developmental speed with size or elaboration at maturity. As such, we consider them an appropriate subject for the extension of theory based on animal study species, while being different enough to be informative in regard to the generality of the hypothesis. For the purposes of this approach, we consider duels and skirmishes fought by the AI to be comparable to various biological scenarios, including actual one-on-one fights between individuals, as featured in the many animal mating systems where rival males clash head-to-head (e.g. [4,33–35]) as opposed to the chaotic multi-combatant fights that occur in other species (e.g. [11,36,37]). In addition, because of the format of the game we consider the AI matches to also serve as models of competition over resources among colonies of individuals, which may similarly unfold in colony-to-colony or multi-colony conflicts, as competition unfolds for example in many species of social Hymenoptera (e.g. [38]) or microbes (e.g. [39]). Those similarities aside, the purpose of this study is to examine armed conflict in an entirely new system. Consequently, we do not expect the AI combatants used here to precisely model any specific system (animal or otherwise). Also note that we did not seek to simulate ecological conditions that might lead animals to fight in duels or not. Instead we experimentally created duel or skirmish scenarios in order to examine the relative advantages of weaponry in these contrasting forms of conflict. To examine our general hypothesis, we specifically tested: (i) whether AI with improved weaponry fared better than expected in duels than in skirmishes, (ii) whether AI that invested more than their competitors in weapon technology were more likely than expected to win matches when facing fewer opponents and (iii) whether AI given an ‘artificial' weaponry advantage were more likely than expected to win matches when facing fewer opponents. We used the RTS war game Starcraft 2 (v. 4.2.0, released 20 February 2018—the most recent version at the time) to set the built-in AI combatants against one-another. Ordinarily, these programmed agents serve as opponents for human players, but in this case we set up matches between the AIs only, without any human players involved. In this game, competitors are simultaneously spawned in separate locations on an arena containing resources, which are harvested and spent in order to buy damage-dealing components referred to as units. The units comprising each combatant are graphically represented as soldiers or fighting creatures, and have the ability to deal damage or use abilities that affect the battle in various ways. Although the units are graphically represented as independent creatures moving in a spatial medium (figure 1), this is for human interpretation only. In fact, the constellation of units controlled by each AI might be better interpreted as the constitutive parts (cells, appendages) available for allocation by the AI at any point in time. Similarly, other dynamic traits of the AI, such as their stockpile of harvested resources, their roster of upgrades achieved and their growth on the arena could be thought of as analogous to forms of animal condition. For example available energy, developmental maturity or body size. However, these are not strict equivalents and we do not seek to treat them as such here. Rather we recognize that, like animals, the AIs have a wide variety of traits that they could allocate resources to, and we seek to test whether weapon elaboration is particularly favoured in duels. We consider the units and their presence in the arena as the AI's equivalent to weaponry because it is through the units and the manoeuvring of them that the AI deal damage to each other. Figure 1. Example of a typical four-way battle, in birds-eye snapshots of the arena taken every 2 min. Black panels indicate the limits of the arena, and coloured dots are units controlled by each of four AI Note that the representation of a swarm is a graphic for human representation only—each unit is intrinsically a part of the AI individual to which it belongs, represented by the four colours. (a) The AI start in opposite corners of the arena. (b–d) The AI use local resources to grow, expanding into the nearby territory. (e) Orange AI makes the first military excursion and (f) attacks the yellow AI unsuccessfully, while the pink and green AI continue to focus on growth. (g–i) The green AI attacks the now weakened orange AI (j–k) The green AI finishes destroying orange, while pink expands territory, attacks and destroys the weakened yellow AI (l–n) The green and pink AI continue to grow, expanding into the territory of their defeated opponents, and exchanging minor battles. (o–p) The green AI makes an unsuccessful attack against pink, losing most of its fighting forces. (q) The pink AI counter-attacks the green peripheral territory, largely destroying it. (r–u) Pink attacks the green core territory, destroying it and consolidating dominance of the arena. (Online version in colour.)
Consequently, the combatant individuals in this system are the AI—they accumulate resources, select strategies, decide when to attack or retreat, develop weaponry (units) and ultimately win or lose the battle by fighting until destroyed. Graphically, the AIs are represented as having an army of units. Mechanically, however, each AI is an individual. It is these AI individuals that we set against one-another for the experiments outlined below. This is similar to animal systems, in which the individual combatants are recognized as such despite being comprised multiple subservient components at different scales, such as appendages, organs, cells, organelles and so on. As such, in our experiments outlined below, duels refer to battles between two AI, and skirmishes to battles with more than two AI. The ability of AI to accumulate new and/or better units follows a branching, hierarchical ‘tech-tree', such that technology investment by the combatant unlocks more advanced units and weaponry over the course of the match. In this way, an arms race is built in to the game design, with a large advantage generally going to the side with superior weapons [40]. The game includes three tech-tree subsets (hereafter named ‘P', ‘T' and ‘Z'), each of which features a unique suite of buildable units and a slightly different method of producing them. A huge and ongoing amount of game design effort is applied by the developer (e.g. [41]) to ensure that these subsets are perfectly balanced (i.e. none has an inherent advantage over either of the others). We increased the combat advantage of a focal AI in two ways; a 20% increase in all unit hit-points (thereby increasing the damage-dealing potential of the AI's weaponry relative to enemies), or increasing a difficulty setting to one higher than all other opponents (focal AI ‘elite' versus ‘very hard' combatants). Increasing unit hit-points represents an improvement to the AI's weaponry in a number of ways, for example:
Increasing an AI's difficulty setting increases its resource collection rate, developmental speed and progression up the tech-tree, making them a stronger fighter more likely to acquire superior weaponry earlier than an opponent AI. We also tested the inverse scenarios with a focal AI given a 20% decrease in unit hit-points or a difficulty setting one lower than all other opponents (focal AI ‘very hard’ versus ‘elite' combatants). We observed battles with the focal AI facing between one and six other AI in free-for-all (FFA) battles (i.e. no allegiances, any contestant able to win by elimination of all others), and tested whether stronger fighters were more favoured in duels relative to skirmishes by comparing the win–loss ratio achieved by a focal AI to the expected null win–loss ratio (i.e. 0.5 in a duel, 0.33 in a three-way battle and so on). We carried out 30 battles under each scenario (tabulated for clarity in electronic supplementary material, table S1), for a total of 720 battles. All battle simulations were carried out on a custom-made, radially symmetrical arena with unlimited resources to fuel the AI Starting positions, AI strategies and AI tech-tree subset selection were randomized in each trial. We considered the type of battle (i.e. duel, three-way and so on) as an ordinal variable and performed a χ2- test for trend in proportions (function prop.trend.test in R, v. 3.3.1, [42]). Because we expected a negative trend in win–loss ratio under the null hypothesis, we artificially modified the argument ‘number of trials’ to equal double the expected number of victories under the null hypothesis (i.e. 15 for duels, 10 for three-way, 7.5 for four-way) instead of being fixed at 30 (e.g. duels: 15 × 2 = 30, three-way: 10 × 2 = 20 and so on). That way for each number of opponents, we had an appropriate null expected number of wins against which to compare the focal AI's observed number of wins. To examine whether weaponry advancement in particular provided an advantage in duels relative to skirmishes, we observed further battles with all AI starting on an equal footing, and compared the weaponry investment of eventual contest winners with that of losers in duels, four-way FFAs and eight-way FFAs. Weaponry investment was extracted from the in-game record of combatant's military research expenditure, referred to as ‘upgrade spending'. We carried out 52 of each battle type, for a total of 156 battles. Arena layout and settings were kept as above, except that no advantage or disadvantage was granted to any AI Post-battle upgrade spending scores for each contestant were recorded as a proportion of the highest upgrade spending score in the game over time. We did not collect data on upgrade spending in the last 60% of the elapsed match time, because in the late-game, combatants which have already gained the upper hand will almost always acquire more resources and invest those in further tech-tree development, creating a feedback between success in the contest and upgrade spending. Because of that, we only wanted to measure early-game investment, and examine how that translated into eventual victory. We fitted a linear mixed-effects model (function lmer in ‘lme4' [43]) with winner relative upgrade spending as the response variable, time (in % of elapsed match time) and battle type (duel, four-way or eight-way) as explanatory variables. We included game identity as a random factor because these measures at different times are not independent from one another (as they are associated with the same game). Type III F-tests using the Satterthwaite's method to estimate degrees of freedom (function anova in ‘lmerTest' [44]) were then used to specifically test the significance of the effects of battle type, elapsed game time and their interaction on winner's upgrade spending. Finally, post hoc tests were used to compare intercepts (function emmeans in ‘emmeans' [45]) and slopes (function emtrends in ‘emmeans') between each battle type. In addition, we observed matches between two to seven combatants, with a focal AI given a selection of high-tech units, while all others were restricted to the three most basic units for their tech-tree subset. By contrast to the previous trials, we did not observe ‘natural' battles with AI following the usual scenario of resource collection and development, culminating in a self-produced set of weaponry. Instead, we set the starting conditions of the match such that each AI began the game with a predetermined, fixed set of military units from the same tech-tree subset, in close proximity to one another. In this way, the experiment was more similar to a classical ethological experiment staged in a laboratory with mature animals in a small arena. This approach, therefore, removed the stochastic influences of resource acquisition, strategy selection, arena exploration and research decision differences between the AI. Instead, they appeared fully formed and adjacent to one-another at the start of the match, and immediately fought to elimination. We repeated this process for each of the three tech-tree subsets, with 30 battles at each number of combatants (between two and seven), thus observing a total of 540 battles. At first we carried out battles with all AI having a number of units that granted them equal total supply (the in-game measure of a unit's value), however, this gave our focal combatant an overwhelming advantage, removing all variation from battle outcomes. We therefore increased the total unit supply of the low-tech AI, to equal 150% of the focal AI unit supply. In this way, the combatants were more balanced for total fighting ability, while the focal AI still had a weapon technology advantage. These contests were carried out on a flat, featureless arena, with opponents starting in contact with each other. The units were arranged as columns (one column for each combatant) in a spoke formation, such that a duel started as a straight line of units, half belonging to each combatant, and a multi-way battle started as a star or asterisk-shaped array, with each arm belonging to a different combatant. Similarly to our first experiment, we compared trend in win–loss ratios achieved by the focal AI as a function of the number of combatants, to the expected null trend using a χ2-test for trend in proportions for each tech-tree. The AIs fought each other to elimination in all cases except for n = 2 draws in the first experiment, which were discarded and re-run. FFA battles were more chaotic than duels, with multi-way skirmishes and flanking manoeuvres commonly occurring (figure 1). AIs that were granted an advantage in either difficulty or unit hit-points relative to their competitors achieved a greater proportion of their null expected victory count in battles with fewer participants (difficulty advantage: χ2 = 16.31, p < 0.0001; hit-point advantage: χ2 = 10.14, p = 0.001), while this trend was reversed when focal AIs were weakened by applying difficulty or hit-point handicaps (difficulty handicap: χ2 = 3.55, p = 0.06; hit-point handicap: χ2 = 2.26, p = 0.13; figure 2). Figure 2. Number of victories achieved by focal AI with an advantage or disadvantage, relative to the expected number under the null hypothesis (i.e. each combatant has the same chances to win), in FFA battles of varying combatant number. The dotted line indicates expectations under the null. (Online version in colour.)
In battles without experimental alterations made to the AI, we found that elapsed game time (F1,1089 = 10.42, p = 0.001), battle type (F2,720.3 = 55.6, p < 0.0001) and their interaction (F2,1089 = 34.6, p < 0.0001) had a significant effect on winner relative upgrade spending. Specifically, duels were won more often by combatants with superior technology in the early game (higher intercept, electronic supplementary material, table S2) relative to their opponents, than were FFAs with four or eight combatants (figure 3). Figure 3. Mean relative weaponry investment of eventual contest winners in duel, four-way and eight-way FFA matches. Elapsed contest time is cut off at 40% due to feedback of contest-winning onto ubiquitous weaponry investment in later stages. Error bars represent standard error of the mean. (Online version in colour.)
Additionally, when focal AI with relatively more advanced weapon technology were set against less advanced competitors, the focal AI won a greater proportion of their expected number of battles in duels relative to multi-way skirmishes. This was the same for each of the three tech-tree subsets (tech-tree ‘P': χ2 = 9.06, p = 0.003; tech-tree ‘T': χ2 = 26.71, p < 0.0001; tech-tree ‘Z': χ2 = 24.61, p < 0.0001, figure 4). Figure 4. Number of victories achieved by focal AI with a technology advantage, relative to the expected number under the null hypothesis (i.e. each combatant has the same chances to win), in FFA battles of varying combatant number. The three different tech-tree subsets in the game are represented by the different line colours. The dotted line indicates expectations under the null. (Online version in colour.)
The dyadic nature of animal contests has long been taken for granted. Models of conflict resolution, staged contests and field observations of species bearing weapons all focus on the one-on-one confrontations of rival males (reviewed by Briffa & Hardy [46]). Although the outcome of multi-player scrambles was known to be less predictable than duels [13,47], the transition to dyadic fighting was never considered to be a potential driver of animal weapon evolution. The importance of duels was recognized first by military historians, who noted that new technologies which brought opponents or vehicles into close contact often sparked an arms race, and that interstate conflicts were most likely to escalate when they coalesced around a pair of evenly matched foes (reviewed by Dupuy [18] and O'Connell [19]). This idea borrowed from the military literature cast animal contests in a new light, and since then the importance of third-party intervention and the possibility of multi-way combat has been gaining attention. In species in which animals live in social groups, but in which conflict arises over resources or social dominance, third-party intervention has been shown to influence coalition formation and dominance hierarchies (spotted hyenas [48]; baboons [49]; ravens [50]). Even in systems that have been thought of as classic examples of duels, such as in contests between male deer during the rut, third-party intervention has been shown to be important [51]. In fallow deer (Dama dama), around 10% of contests between stags are disrupted by interveners [51] and interveners improve their mating success [52], while those suffering intervention, have reduced mating success [53]. Clearly, not all animal contests are duels, and the proportion of contests occurring as duels may help explain population and species differences in relative weapon size (e.g. [13,54]). If true, then the behavioural transition from scrambles to duels could underlie the initiation of arms races in both animal and military weapons, highlighting a simple property of confrontations with far-reaching consequences for all types of conflict. Here, we tested whether dyadic encounters are important for weapon evolution using a system analogous to, yet nonetheless distinct from, both animals and military technologies; cybernetic combatants in computer war games. We found that increasing numbers of cybernetic combatants reduced the benefits of weaponry or power that occur in dyadic contests. In duels, combatants with greater power (as granted in two different forms) had a large advantage, but that advantage deteriorated as the number of competitors increased. Conversely, experimentally weakened contestants suffered less of a disadvantage in multi-way skirmishes, winning relatively more often in contests with more combatants. In particular, combatants with the highest early-game investment in weaponry often went on to win duels, but in FFAs, this trend was reduced. Furthermore, when we experimentally granted focal AIs a weaponry advantage in the form of high-tech units, they enjoyed a greater benefit in duels relative to multi-way skirmishes. Therefore, our results suggest that the evolutionary hypothesis regarding the role of duels in animal and military weapon evolution may accurately reflect underlying natural laws of conflict, and possibly explain the occurrence of arms races in disparate duel-like systems. It is interesting that strengthened AIs won fewer than their null expected number of victories in skirmishes with more than four combatants, because it suggests that in contests against multiple combatants, their greater power actually became a disadvantage in some way. This could be due to early-game dominance granting them large territories which became hard to manage at later stages, or through hyper aggressive strategies causing them to ‘burn out' against multiple opponents while more passive tactics endured longer. Assessment of these possibilities was beyond the scope of the current study, but they open interesting questions for future examination. It may be that similar feedbacks restrict strategies based on high aggression or weapon specialization in biological systems where interactions between more individuals at any given time are common. Animals with exaggerated weapons are known to suffer metabolic, biomechanical and locomotory costs [55–59]. Likewise, there is growing evidence that weapons show developmental trade-offs with testes [60,61] and other morphological structures that grow in proximity to weapons [62]. Exaggerated weapons that convey an advantage in one context (fights) have also been shown to be a disadvantage in another context (races through tunnels) in the dung beetle Onthophagus taurus [14]. Here, we have shown that although the exaggerated weapons of AIs conveyed an advantage in duels, that advantage deteriorated in multi-opponent skirmishes. In the context of our second experiment, this could be due to the AI's investment in weapon technology trading off with alternative strategies that may increase their chances of victory against multiple opponents. This trade-off may therefore favour different resource allocation strategies depending on the number of opponents. Our third experiment also suggested that another trade-off may occur entirely within the context of fights themselves as the advantage of superior weaponry declined with greater numbers of opponents, even with the potential for alternative strategies (such as rushing or sneaking) removed. In that experiment the only variable other than weaponry was the number of opponents, so it could be that choices such as how many different opponents to engage at a time, or who among them to engage, might overshadow weaponry differences at high contestant numbers. We hope that studies of animal contests between varying numbers of competitors might evaluate whether such a trade-off also occurs in nature. Because our results arose in a non-biological system, they suggest that duels may feed arms races not only in the context of exaggerated weaponry and animal contests but also in other systems as well. These might include evolutionary arms races such as those between parasites and hosts, or arms races in non-biological systems such as human technology, business, military escalation, trade wars or cyber warfare. In many of these scenarios, conflicts can be played out in a duel-like fashion (such as between a specialist parasite or predator and its crucial host or prey species) or more skirmish-like (such as between generalists). Indeed, this logic tracks closely with the idea of ‘hot spots’ and ‘cold spots’ in the geographical mosaic theory of coevolution [63]. In any scenario where the only way to win (or survive, or reproduce) is to outcompete a single rival that is in the same situation, we might expect extreme development of something analogous to weaponry or resource holding potential. Although game theory has been critical to explaining various evolutionary phenomena (e.g. [64]), the use of actual games to test evolutionary theory has not been employed to our knowledge, beyond the use of pure mathematical models. Our results highlight the potential to examine complex biological contests with the pre-existing conflict simulators that are provided in the form of consumer war games. In this way, other aspects of contests could also be examined. Various evolutionary and ecological concepts have equivalents in war gaming—resident versus intruder status in animal contest research [65] is analogous to the ‘defender's advantage' in war gaming, assessment of rivals [12] occurs during ‘scouting' in war gaming and rock–paper–scissors-style alternative strategies (e.g. [66]) are paralleled in the concept of ‘build-order-wins'. Fighting technique also has an effect on the type of weaponry that evolves [67,68], and war games also usually feature a range of possible techniques that can be employed, some focused on quickly acquiring certain weapon technology, and some focused on making efficient use of low-tech units. Another area which is beginning to be incorporated into evolutionary models of contests is the potential for offense to result in damage to self as well as damage to the opponent [69,70]. This is also something that could be easily examined with war game simulations, as aggressive forces suffer clear losses while attacking (in fact, these data are specifically quantified and reported by software designed to aid match analysis and commentary). Finding that the predictions of hypotheses can apply to diverse systems suggests a match between the underlying theory and the laws of nature. Thus, we should seek to test predictions in systems different to those for which they were developed. We have done so with the current study, examining a hypothesis relating to the evolution of animal weaponry in a completely separate context. We ultimately found that in contests between AIs in a computer war game, the advantages of exaggerated weaponry in duels were negated in multi-opponent battles. We are unaware of studies of animal contests that specifically test whether third-party interventions negate or reduce any benefits of larger weaponry or other measures of resource holding potential, but our results suggest that they would, and that this could be a fruitful area for future research. This article has no additional data. We declare we have no competing interests. We acknowledge funding from NSF grant no. IOS-1456133. FootnotesElectronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.5001068. References
Page 13Colonies of social insects such as ants have been likened to multicellular organisms (‘superorganisms’) for well over a century [1–3]. This analogy rests on the definition of a superorganism as a society with irreversible morphological specialization and mutual interdependence of reproductive and non-reproductive castes (for a review on the history of the usage and interpretation of the term, see [4]). Queens and males are the reproductive tissues of the colony organism, and workers the somatic tissue, sometimes further diverged into different subcastes, specialized into defence, foraging and brood care, for example. The evolution of such ‘superorganismal’ societies is a key example of a major transition in evolution [4–6]. Despite the popularity of the superorganism concept, the extent to which it has directly motivated empirical or theoretical research is arguably limited (but see [7] for a recent review). There are insightful descriptions of colony function based on hierarchical comparisons of the anatomy, physiology and regulation in multicellular organisms and insect societies [8,9]. However, no novel testable predictions have been emerging from such approaches, and despite the extremely diverse life histories and allocation strategies of superorganisms [10], few attempts to explain the variation with organismal, rather than social insect specific theories, have been made (but see [11] for a notable exception). Using the organismal analogy to explain life history and allocation strategies of social insect colonies [10] requires adapting existing theoretical models to fit social insect biology, as their assumptions will typically not directly fit the colonial superorganism. In this paper, we make such an attempt and apply gamete competition models of anisogamy evolution (i.e. the divergence of gamete sizes, such as egg and sperm [12–15]) to superorganisms, motivated by the verbal analogy of Helanterä [10]. If we take a social insect colony as a superorganism rather than a society formed of individual insects, the dispersing young males and future queens are analogous to the gametes of a multicellular organism, i.e. sperm/pollen and eggs/seeds, respectively. They leave their mother colony, disperse and mate to form a zygote-like incipient colony, i.e. a mated queen (accompanied by a king in the case of termites) ready to lay eggs that will develop into workers, that form the somatic body of the colony. This ‘zygotic’ stage of obligatory production of non-reproductive workers is what sets apart a superorganismal life cycle from a female solitary organism who can reproduce immediately after mating. That is, there is an obligatory somatic growth stage without which reproduction is completely impossible [16] (figure 1), analogous to the development of a zygote where all fitness gains are lost if the zygote does not survive into the reproductive stage of the life cycle. Figure 1. Gametic reproduction results in fitness benefits if the zygote survives. Reproduction by a queen founding a superorganismal colony results in fitness benefits (i.e. queens and males that mate to form the next generation) if the incipient colony survives, analogous to the gamete-level case. By contrast, in reproduction by solitary organisms, there is no stage in the life cycle that is analogous to a zygote or incipient colony, and a mated female can immediately reproduce. This is why there is a clear analogy between the evolution of gamete dimorphism and queen–male dimorphism in eusocial insects, but not between the evolution of gamete dimorphism and sexual dimorphism in solitary organisms.
The gamete competition theory for the evolution of anisogamy aims to explain the divergence of gamete sizes into male and female gametes, and analyses the conditions under which such a divergence is expected to occur. The model assumes that there is ‘fair raffle’ type of competition between gametes over fertilizations, resulting in a selection pressure to produce more numerous (and hence smaller) gametes. At the same time, the developmental requirements of the zygote result in a selection pressure to produce larger gametes for provisioning the zygote. In its modern form [17], gamete competition theory suggests that gamete sizes begin to diverge once the developmental requirements of the zygote diverge far enough from the minimum size requirements of the gamete itself, as is expected to happen as multicellularity evolves. Under these conditions, disruptive selection can lead to the production of large gametes (e.g. eggs) that provide resources to the offspring, and small gametes (e.g. sperm) that ‘compete’ for fertilizations. Gamete competition theory remains the most widely accepted explanation for the evolution of anisogamy, and its predictions have empirical support, although exceptions are also known (see [13–15] for recent reviews). Analogues of the basic assumptions of gamete competition models are likely to apply at the superorganism level as well. First, even if male behaviour, queen mate choice and sexual selection in social insects are still understudied (but see [18,19]), the assumptions of random mating with respect to size and a scramble-like competition for matings (analogous to fair raffle competition at the gamete level) are likely to largely apply [10]. There is little evidence of aggressive male–male competition or mate choice by queens [20]. Second, resources available to the ‘zygotic’ incipient colony increase its survival prospects, so that initial colony (zygote) size brings benefits of size beyond survival to mating. The more resources there are available at the very earliest stage of colony founding, the more workers can be produced rapidly without external resources. The colony foundation stage when the colony consists of a queen and few workers only is a period of high mortality due to conspecific competition [21,22], such as brood raiding in the fire ant Solenopsis invicta [23]. If fast growth is indeed a key determinant of colony survival, and if a queen with ample resources can lay eggs faster, the assumption of a positive relationship between initial colony resources and colony survival (analogous to a positive relationship between zygote size and zygote survival) probably holds. One key point where the process differs at the gamete and superorganism levels is that in some species of ants, colonies are founded by multiple queens [24,25] and this may alter the queen size–colony survival relationship (see model 3). Thus, superficially the analogous evolution of anisogamy at two different hierarchical level seems a feasible hypothesis. Furthermore, even if the empirical patterns have not been systematically analysed, at least in ants, queens seem to be consistently larger than males (see [10] and references therein). In this paper, we adapt gamete competition models of anisogamy evolution to fit the reproductive biology of superorganisms. We explore whether similar processes could underlie evolution of anisogamy at different hierarchical levels and highlight the key similarities and differences in the model predictions. Our models are influenced by several earlier models on the evolution of anisogamy, but the central paper that we have followed is that of Bulmer & Parker [17], a game theoretical update of classic theory on the evolution of anisogamy [12]. The work of Bulmer & Parker is a particularly suitable starting point for us for several reasons. It begins with the assumption of pre-existing mating types (contra [12], in which all possible pairs of gametes were able to fuse with each other), analogous to our assumption of pre-existing male and female sexes. Second, it analyses the requirements for the evolution of anisogamy and how they relate to size-specific survivorship of gametes and zygotes: one of our aims is to determine whether an equivalent of these requirements exists at the superorganism level. Third, it is arguably the most complete analysis of gamete competition models while excluding gamete limitation, which is an alternative (or complementary) pathway to the evolution of anisogamy [13]. Later analyses of anisogamy evolution have combined both gamete competition and gamete limitation in a single model (beginning with [26]), but here we exclude gamete limitation and its superorganismal equivalent mate limitation for biological reasons (it is reasonable to assume that matings of the queens are not limited by availability of males) as well as reasons of clarity and mathematical simplicity. Thus, our primary aim is to determine whether an equivalent of gamete competition models of anisogamy evolution can drive the evolution of queen–male dimorphism in superorganismal insects. We begin with a modification of the first model (model b in their notation) from Bulmer & Parker [17]. We focus on the salient differences between the evolution of gamete dimorphism versus queen–male dimorphism, and do not explicitly consider the role of workers or any relatedness asymmetries. We simply assume a fixed sex allocation of 1 : 1, that the size of queen and male offspring are evolving traits and that there is a positive relationship between queen size and colony survival. Notation and definitions for all models are given in table 1. The functions for offspring size-number trade-offs, for individual survival and for colony survival in table 1 follow those used by Bulmer & Parker [17] in their gamete-level model. The function used for colony survival was originally derived for survivorship of marine invertebrate eggs [27,28]. It is based on the premise that larger eggs develop more quickly and are thus under mortality risk for a shorter period of time. The derivation starting from this premise [27,28] results in the function s(z) (table 1), which increases initially in an accelerating fashion until the increase eventually slows down and the function approaches a maximum value of 1 (corresponding to a maximum survival probability of 1). The parallel premise of decreased colony development time with larger queen size is reasonable for superorganisms [10], hence justifying the use of the same function here. It is not as obvious that the same function should apply to the survival of individual queens and males, and hence we follow Bulmer & Parker [17] and use the same function (denoted g(z) in table 1) for individual survival (model 1), but also take the alternative approach of having a fixed minimum size (model 2).
Total allocation to male and queen offspring is assumed to be equal, but offspring numbers depend on their sizes. We assume that all queens are fertilized on their nuptial flight, so that males are effectively competing for a fixed total fitness pot for any given queen size. As a clear departure from the logic of the anisogamy model of Bulmer & Parker [17], in the current model, males only contribute genes and no material resources for colony or offspring development (as is the case in social Hymenoptera, where males die after mating without further contributions to colony success). In gamete evolution models, the size of both gametes can contribute material resources to offspring fitness, particularly when the two gametes are similar in size. A spermatozoon is typically assumed to contribute near-zero resources, but tiny sperm represent only one end of a continuum from isogamy (equal contribution) to extreme anisogamy. This difference removes one coevolutionary aspect of the anisogamy model. Males face scramble competition so that each surviving male has equal chances of mating regardless of his size. These assumptions imply that the structure of the model is unaffected by the number of mating partners of a queen or male: if each male is considered analogous with a raffle ticket, then the expected winnings per male do not depend on the extent of multiple mating by the queens (although the distribution and variance of male fitness outcomes may do so). In other words, if total fitness is wtot and there are N males competing for it, then the mean fitness per male is wtot/N so that the expected value of each raffle ticket is the same regardless of how the prize is distributed between them. We assume that the male population is large and well mixed to the extent that when a mutant mother producing a slightly deviant number of male offspring appears, this initially has a negligible effect on the extent of competition faced by male offspring. In other words, although a mutant queen produces n^y sons, there are still on average ny males competing for every nx queens in the population, and in the competitive environment of any male seeking matings. We also assume that mutations affecting male and queen sizes appear rarely enough that the two types of mutations are never segregating in the population simultaneously. Therefore, mutant males never compete for mutant queens. Now we can write the fitness of a daughter carrying a mutant allele as the product of her own survival until colony founding, and the subsequent survival of the colony g(x^)s(x^).2.1 The fitness of a son carrying a mutant allele is g(y^) nxg(x)s(x)nyg(y).2.2 The logic of equation (2.2) is as follows: in the well-mixed population, an average of nxg(x)s(x) colony fitness units are distributed over nyg(y) surviving males. Hence, the latter component of equation (2.2) is simply the average amount of fitness per male surviving to mate. The focal mutant male survives to mate with probability g(y^), yielding an average fitness per mutant male of g(y^) nxg(x)s(x)nyg(y). Hence, the total fitness a focal mother gains via all mutant daughters is w^x=n^xg(x^)s(x^),2.3 and via all mutant sonsw^y=n^yg(y^) nxg(x)s(x)nyg(y).2.4 Substituting the functions for n, g and s (table 1), we can compute the direction of selection for each as 1wx∂w^x∂x^|x^=x=(n^x′nx+g′(x^)g(x) +s′(x^)s(x))|x^=x=−1x+αx2+βx22.5 and1wy∂w^y∂y^|y^=y=(n^y′ny+g′(y^)g(y) )|y^=y=−1y+αy2.2.6 The method used in equations (2.5) and (2.6) is that of evolutionary game theory [29], which is in the context of this model effectively equivalent to adaptive dynamics (i.e. the equations are of similar form in both methodologies; see [30]). In words, equations (2.5) and (2.6) amount to estimating selection on mutants that deviate from the prevalent resident size by a small amount. Clearly, there is an asymmetry in these equations that is not present in a typical anisogamy model (e.g. [17]). The equations of selection in anisogamy models are typically symmetrical, and either gamete type can become the larger or smaller one. In equations (2.5) and (2.6), selection on queen size includes an additional component β/x2 that is missing from selection on male size, and it is thus predetermined that queens become larger than males. The candidate evolutionarily stable strategies (x∗ and y∗) for male and queen size can be solved by setting both equations equal to zero and solving for x and y (i.e. finding values for queen and male sizes where selection vanishes). A straightforward calculation yields x∗=α+β2.7 andy∗=α.2.8 These equations show that under the model assumptions, queens are expected to be larger than males. Stability analysis (electronic supplementary material, appendix) shows that x∗ and y∗ are convergence stable and evolutionarily stable (see ch. 12 of [31]). It is worth elaborating once more on why the result is different from that of the anisogamy model of Bulmer & Parker [17]. The reason is that in the anisogamy model, both gametes have potential to provision the zygote and thus influence its survival, and the resource provisioning by both gametes can be substantial in isogamous and near-isogamous reproductive systems. In the current model, males play no part in colony survival after fertilizing a queen. As the male gamete size does not enter the zygote's survival function, there is no feedback between male and queen sizes, and hence no coevolution. Mathematically speaking, where ‘zygote’ fitness in the current model is s(x), in an anisogamy model, it would be s(x+y), hence linking the fitnesses of the two gamete types in a relatively more complex coevolutionary process. In the current model, both classes reach their evolutionarily stable sizes independently. For principally the same reason, sizes of males and queens always diverge in the current model as long as β>0, whereas in the anisogamy model of Bulmer & Parker [17], they only diverge under the condition β>4α (that is, when the survival requirements of the zygote have sufficiently diverged from those of the gamete, as is expected when multicellularity develops). Now the analysis is repeated adapting model d from Bulmer & Parker [17]. In their model d, the authors assume that gamete survival is dependent on size in a stepwise fashion, so that below a threshold size δ, gamete survival equals 0, and above the limit, it equals 1. In our colony-level model, this implies that fitness via both queen and male offspring is 0 below that limit, while above the limit, they are s(x^)2.9 for queen offspring andnxs(x)ny.2.10 for male offspring. Equations (2.9) and (2.10) are equivalent to equations (2.1) and (2.2) with the gamete survival functions (g) set equal to 1.For viable sizes (>δ), total fitness a focal mother gains via all mutant daughters is w^x=n^xs(x^)2.11 and via mutant sonsw^y=n^ynxs(x)ny.2.12 Substituting n and s (table 1), we can compute the direction of selection for each as 1wx∂w^x∂x^|x^=x=(n^x′nx+s′(x^)s(x))|x^=x=−1x+βx22.13 and1wy∂w^y∂y^|y^=y=(n^y′ny )|y^=y=−1y2.14 Again the outcome is clear. For male offspring, selection is always negative above minimum size δ, and hence they are driven towards this minimum by selection. For queens, a candidate ESS is found at −1x+βx2=0, or x=β.2.15 Hence, for model 2, as long as β>δ, the queen candidate ESS is larger than male candidate ESS. A stability analysis again confirms the male and queen equilibria are both evolutionarily and convergence stable (electronic supplementary material, appendix). Now we consider a situation that does not have an analogue in the gamete evolution models of Bulmer & Parker [17]: multiple queens founding a colony. We modify model 1 (above) for this purpose. The g-functions are not altered, but consider the s-function. If colony survival depends on foundress resources as s(z), then in this case z will be composed of resources from multiple queens. Let us for simplicity assume that s depends simply on the sum of resources from all founding queens, and the functional relationship s(z) remains as in table 1. Hence, if k queens in total contribute to a colony, of which one is a mutant, then colony survival is s(z)=s(x^+(k−1)x)2.16 where x^ is mutant queen size and x resident queen size. Total mutant fitness via queen offspring is thenw^x=1kn^xg(x^)s((x^+(k−1)x)).2.17 Where we have assumed that on average colony fitness is split evenly between foundresses. This could correspond to, for example, equal sharing of the colony, or to only a single, randomly picked foundress surviving to ultimately reap all the benefits of the collaboratively founded colony, as usually is the case [25]. We therefore assume that although the resources an individual queen brings do alter survival prospects of the colony, they do not alter her expected relative share of fitness from a successful colony.Total fitness via mutant sons w^y=n^yg(y^)(1/k)nxg(x)s(kx) nyg(y).2.18 The candidate ESS for male offspring size is as in model 1, but for queen size, the direction of selection is now determined by 1wx∂w^x∂x^|x^=x=(n^x′nx+g′(x^)g(x) +s′(x^+(k−1)x)s(kx))|x^=x=−1x+αx2+βk2x2,2.19 and the candidate ESS is correspondinglyx= α+βk2.2.20 So, under the multiple foundress model, queens are still expected to be larger than males, but the size difference rapidly decreases with the number of queens. When there is only one queen (k = 1), model 3 is identical to model 1. Again, the results are convergence stable and evolutionarily stable when k = 1 or k = 2. For larger values of k, stability analysis is more subtle, but for a very wide range of biologically realistic parameter values, the equilibria remain stable (details are given in the electronic supplementary material, appendix). Our analysis raises a number of interesting points and may clarify understanding of both the evolution of queen–male dimorphism and that of anisogamy. First and foremost, we have confirmed the verbal idea [10] that the evolution of queen–male dimorphism can, at least in principle, evolve via similar logic as gamete dimorphism in classic models of the evolution of anisogamy (e.g. [12,17]). Compared with solitary lifestyles, the superorganism life cycle includes the additional stage of colony founding, which in turn introduces a new pressure: all fitness of the queen is dependent on survival of the colony that she founds until the stage where new queens and males are produced, and this colony-level reproduction is impossible before the ‘somatic’ colony has reached a certain size, such as a threshold number of workers [32]. By implication, if colony survival at the founding stage is dependent on the size of the queen, there will be a period of selection on queen size that is absent in the life cycle of a solitary insect. This is the reason we can ‘frame-shift’ upwards, and draw an analogy between queens, males and the superorganismal colony on the one hand, and sperm, ova and the organismal zygote on the other. Such an analogy does not exist in solitary insects, because there is no equivalent of the zygote if we introduce a similar frame-shift (figure 1). Despite the apparent analogy, we have also clarified a fundamental difference between anisogamy evolution and the ‘superorganismal anisogamy’ modelled here. In the current model, although male fitness is dependent on the survival of the colony, males do not contribute in any material way to the colony: they contribute only their genes, no matter what their size, in contrast with the evolution of gamete sizes where each gamete contributes resources according to size. This assumes that no resources that facilitate productivity are transmitted in the seminal fluid, which is currently unknown in social insects. Males therefore have no influence on the survival prospects of the colony. This is in stark contrast to models of the evolution of anisogamy, where two gametes fuse to combine not only genetic contributions, but their material contributions too (particularly when the two gametes are similar in size). It is this combination of material resources and the gametes' mutual influence on the survival prospects of the zygote that make the anisogamy model fundamentally coevolutionary in nature. It is also the reason why the necessary conditions for the divergence of gamete sizes in the anisogamy model are more stringent than the conditions for divergence of queen and male sizes in the current model. In the anisogamy model, the energy requirements of zygotes must be sufficiently larger than those of gametes to allow the divergence of gamete sizes [17]. In the current model, there is no such requirement. Even if colony survival places a fairly minor additional burden on the queen, queen–male dimorphism is expected to evolve. A second difference is that in some superorganismal species, the colony is founded by more than one queen, whereas a zygote is founded by one ovum only: we have investigated the implications of this difference in model 3. Third, an ovum is fertilized by one sperm only (polyspermy typically leads to abnormal development [33,34]), and conversely, a spermatozoon fertilizes only one ovum. In superorganisms, neither is necessarily true. A queen can (and often does [35]) mate with multiple males, and a single male can potentially mate with several queens [20]. Neither of these differences alter our model: if mating is random and if all queens are fertilized, the average fitness gains per male remain the same regardless of these differences. Finally, in contrast with a female gamete of a normal organism which ceases to exist as an independent entity from the zygote stage onwards, the social insect queen remains a physically independent unit throughout the colony life, so the link between queen size and colony reproductive output is not necessarily completely severed. This would potentially further increase dimorphism, but the effect should be similar or smaller in superorganismal species relative to solitary species, because the relationship between queen size and reproductive output is decoupled to some extent by ‘somatic’ workers. The key differences between the fitness effects of solitary female and superorganismal queens require some clarification. Selection for larger female size may, of course, well occur in a solitary species as well, both due to size-related increase in survival until mating, in further survival until reproduction, and fecundity. Thus, it might be challenging to tease apart whether precisely the ‘anisogamy-like’ processes would be driving queen–male dimorphism in social insects (if such a pattern does indeed hold up to systematic scrutiny of broad, phylogenetically controlled patterns). However, it is worth pointing out that (i) a size-related increase in survival until mating is unlikely to be systematically female-biased, (ii) size-related increase in survival between mating and reproduction is unlikely to apply in a majority of solitary species (i.e. outside species that do not construct nests, gather resources, defend a territory etc.). Furthermore, it is possible that large size provides fecundity benefits to a superorganismal queen as well, although indirectly, as the reproductive output is mediated by the ‘somatic’ worker phenotype of the colony. The key point remains: the obligatory growth stage of worker rearing after mating is a stage where positive selection for size occurs in queens but not males of superorganisms, but an analogous stage is lacking, or at least is not always present in solitary insects. The model suggests several comparative tests [10]. The primary prediction is that dimorphism should be higher in superorganismal than solitary species, or species with more primitively social lifestyles. Second, model 3 shows that superorganismal species where multiple queens found a colony are expected to have lower size dimorphism than those where colonies have only a single foundress, and dimorphism is expected to decrease with the number of foundresses. Equation (2.20) suggests that size dimorphism should decrease quite quickly with the number of foundresses (i.e. in inverse proportion to the square of the foundress number, although the exact mathematical form will depend on model assumptions). The reason for the fast decrease is that each queen is able to gain from the resources brought by other queens, which allows ‘exploitation’ of the shared resource with a reduced contribution of one's own. These pattern may change, however, if queen size increases her chances of being the sole survivor until colony maturity (e.g. [36]). Third, in species where the assumptions of the model are not met, dimorphism should be smaller. For example, effects of male size on mating or mate finding success may vary according to male life histories [37]. Selection pressures on queen size may also have been relaxed in species where workers from the mother colony accompany the foundress queen in so-called dependent colony founding through, for example, swarming or foundation of bud nests [38]. Dependent colony founding is the prevalent mode in Apis bees, and has evolved repeatedly in ants as well. Also differences in colony founding behaviour of single founding queens may introduce variation into selection for size. The benefits of size might be larger in queen in the so-called claustral founding species where the queen does not forage, but depends solely on her internal resources for rearing the first workers. Finally, biases in dispersal abilities of queens and males, and more generally lack of panmixia may affect selection for size (see [26] for a model of anisogamy evolution in a population where gamete dispersal is limited). While social insects with large-scale mating flights are often assumed to be panmictic [10], male-biased dispersal and limited gene flow through females is also occasionally reported, especially in ants [39]. However, some degree of male-biased dispersal by itself does not necessarily breach the assumptions of our model: the central assumption is that males disperse to a sufficient extent that competition between sons of a rare mutant queen is negligible. Fourth, while our model is based on a social Hymenoptera life cycle where the male dies soon after mating and does not contribute to colony life, an interesting comparison is provided by termites where both sexes stay in the colonies and contribute to the colony foundation and brood care. Thus, termites are more similar to the original anisogamy model, so the conditions for divergent selection are potentially more stringent and depending on a coevolutionary process. It should be noted, however, that a termite-specific model would still not be completely symmetrical as anisogamy models are: even if males contribute to colony founding, their contribution is of a fundamentally different kind than that of queens who are able to reproduce. There are reported cases where termite queens are slightly larger than the males [40,41], but broad comparisons across species are lacking. A further complication for testing the predictions may arise from selection for larger males due to male–male competition [42]. Models of anisogamy evolution under gamete competition make testable predictions, some of which apply to the current model. A prediction that is transferrable to the current model is that gamete dimorphism is predicted to increase with increasing size and complexity of the adult organism [12,17,43], and this prediction has some empirical support from comparative studies [44–46]. In the superorganism model, the analogous prediction is that queen–male dimorphism is predicted to increase with increasing size (i.e. number of workers in a colony at sexual maturity) and complexity of the colony. While colony complexity might be difficult to quantify unambiguously, reasonable proxies include, for example, the presence and number of worker subcastes, or number of workers. In the model, the reason for this prediction is that a larger and more complex superorganism probably requires a larger amount of initial resources for reasonable chances of survival, so that we should expect a positive correlation between size/complexity and the parameter β in the model. This assumed correlation again goes back to an analogy with models of anisogamy evolution. In the gamete competition model of anisogamy evolution, the parameter β in the zygote survival function s increases with the need to provision the embryo, and the needs of the embryo are in turn thought to increase with increasing multicellular complexity of the adult organism [17]. Analogously, the parameter β in the colony survival function s in the current model is thought to correlate with the complexity of the colony. A prediction that is not transferrable is that anisogamy should only evolve if there is a sufficiently large difference between the minimum requirements of the gamete and zygote [17], which also suggests that there should typically be a near-stepwise distribution from isogamy with small gametes to anisogamy with a relatively large difference in size between the male and female gamete [47] (this prediction has some empirical support [47]). This prediction is not replicated in the superorganism model because the coevolutionary link between queen and male size is broken (see model 1 description). Hence, the current models predict a relatively continuous distribution of queen–male dimorphism in superorganisms, compared to the expected stepwise distribution of gamete dimorphism from isogamy to anisogamy. To conclude, we have shown how adjusting a classic model of anisogamy evolution to the biological particularities of social insect colonies as superorganisms can elucidate the evolution of size dimorphism among social insect queens and males. Thus, the superorganism metaphor can be used as a starting point to derive testable predictions, and ultimately increase our understanding of insect societies, building on theoretical and empirical understanding of how multicellular organisms evolve. Furthermore, the model highlights similarities and crucial differences of evolutionary processes at different hierarchical levels, and thus increases our understanding of the major transitions in evolution. This article has no additional data. Both authors contributed to all aspects of the manuscript. We declare we have no competing interests. J.L. is funded by an Australian Research Council Discovery Early Career Research Award (project no. DE180100526) from the Australian Government. H.H. has been funded by the Kone Foundation. We are grateful to Jacobus J. Boomsma, Madeleine Beekman, Geoff A. Parker and two anonymous referees for helpful comments on earlier drafts of the manuscript. FootnotesElectronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.5004641. © 2020 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original author and source are credited. References
Page 14Prior to mating, animals advertise and evaluate an array of often conspicuous visual, acoustic and chemical sexual signals [1,2]. These pre-mating sexual signals offer the choosing sex (typically females) the opportunity to assess the quality and genetic compatibility of potential mates [1–4]. After mating, communication between the sexes continues but is restricted to chemosensory communication between gametes and, in the case of internal fertilizers, the female reproductive tract [5]. Such post-mating chemosensory communication between eggs and sperm can facilitate gamete-mediated mate choice, allowing eggs to exert cryptic female choice and bias fertilizations towards specific males [6,7]. However, our understanding of the potential for gamete-mediated mate choice remains limited in internally fertilizing species, and is completely unexplored in humans. Sperm chemoattraction, a remote form of chemical communication between eggs and sperm occurring before gamete contact, is a widespread mechanism for increasing sperm density around unfertilized eggs in animals [5]. In broadcast spawning marine invertebrates, where adults are unable to express pre-ejaculatory mate choice, chemoattractants increase fertilization rates by increasing the effective target size of the egg, maintain species barriers by preferentially recruiting conspecific sperm, and allow sperm and eggs to exercise gamete-mediated mate choice [6–9]. Chemoattractants in marine invertebrates can also preferentially recruit sperm from specific, presumably more compatible males by remotely altering sperm swimming physiology and behaviour, thereby increasing fertilization rates, embryo viability and offspring survival [7]. In internally fertilizing species, females can exercise cryptic female choice through interactions between sperm and the female reproductive tract, influencing the number of sperm a female retains and/or sperm swimming performance [10,11]. Subsequent interactions between eggs and sperm can also facilitate gamete-mediated mate choice in internal fertilizers. For example, in house mice (Mus domesticus) Firman & Simmons [12] found that eggs were preferentially fertilized by sperm from less related males during in vitro fertilizations (IVF), and suggested that either direct interactions among cell-surface proteins on gametes or differential responses to chemoattractants could explain these effects. In mammalian reproduction, chemoattraction is the last of a series of sperm guidance mechanisms (including positive rheotaxis and thermotaxis) that acts to recruit capacitated sperm to eggs [5,13]. By contrast with marine invertebrates, mammalian sperm lack species-specificity in responses to chemoattractants [14], suggesting that pre-mating species recognition mechanisms may reduce the need for post-mating processes to reinforce species barriers. Nevertheless, mammalian chemoattractants could play a post-mating role in gamete-mediated mate choice, either to maximize genomic compatibility between potential mates [3,4] or to reinforce or override pre-mating mate choice decisions [11,15]. However, the potential for chemoattractants to serve a sexually selected role in human reproduction remains unexplored. Here, we assess if follicular fluid, a source of sperm chemoattractants [16], differentially regulates sperm behaviour to reinforce pre-mating mate choice decisions and mediate fertilization success in humans. Human sperm respond to chemoattractants present in the follicular fluid surrounding eggs (most likely progesterone [5], although this remains a source of ongoing debate) by altering their swimming behaviour to orient towards, and accumulate in, follicular fluid [16]. Sperm behavioural responses can differ among follicular fluids, such that follicular fluids from different females exhibit variation in their ability to attract sperm from the same male [16]. Moreover, females producing follicular fluid that was better at causing an accumulation response in sperm also produce eggs that achieved higher fertilization rates in clinical IVF cycles [16]. Thus, differential responses in sperm behaviour to follicular fluid have the potential to facilitate gamete-mediated mate choice in humans. We investigated this potential using two distinct experimental designs, exposing sperm to follicular fluid from two females either simultaneously or non-simultaneously, and report robust evidence that sperm accumulation is influenced by the interactive effects between males and females. We obtained follicular fluid and sperm samples from couples undergoing assisted reproductive treatment (IVF; intracytoplasmic sperm injection, ICSI) at St Mary's Hospital, Manchester, UK, with written informed patient consent and approval from Central Manchester Research Ethics Committee (electronic supplementary material). We specifically focused on couples receiving assisted reproductive treatment, rather than, for example, performing assays using sperm from males not seeking fertility treatment, as one of our aims was to investigate the link between partner choice and gametic interactions. Samples were obtained using standard clinical practices [17] (see electronic supplementary material). All data were collected blind to the treatments and patient identity to ensure that patient confidentiality was maintained to comply with the WHO Good Clinical Research Practice guidelines [18] and the Human Fertilization and Embryology Authority Code of Practice [19], and also ensure that the researcher was blinded from identifying which samples originated from each couple during the experiment. Information relating to the participant's fertility procedure was collected, including the type of fertilization method used (IVF or ICSI), number of oocytes retrieved, number of oocytes successfully fertilized, embryo quality score, pregnancy outcome and live birth success. For couples undergoing IVF, fertilization success was calculated as the number of fertilized embryos divided by the number of oocytes retrieved and inseminated, while for couples undergoing ICSI, fertilization success was calculated as the number of fertilized embryos divided by the number of oocytes injected. Embryo quality was determined using an embryology morphology grading scheme (see electronic supplementary material). Pregnancy outcome was scored based on evidence of implanted embryos, while live births were treated as successful outcomes relative to all other outcomes (see electronic supplementary material). Ejaculate traits differ between patients being treated with IVF or ICSI; for example, sperm density is lower in ICSI than IVF patients in the non-simultaneous choice experiment (see below) (linear mixed model, LMM: χ2 = 5.2, p = 0.02). Therefore, we were cautious in how we analysed data from these two treatment groups. However, our experimental design excluded males with severe male factor infertility, as such males would not have sufficient sperm to be used in the experimental assays (i.e. males undergoing ICSI were diagnosed with either male factor subfertility or were normospermic men who suffered poor fertilization with IVF in a previous cycle). Data analyses examined the impact of including ICSI patients in statistical models either by removing these patients or including the type of fertility treatment as a fixed effect in the model (described below). To test if follicular fluid influences sperm behaviour, we adapted a classic dichotomous mate choice assay to the microscopic scale (i.e. simultaneous presentation of two stimuli; figure 1). We performed two experiments using a North Carolina II cross-classified block design [20]. This experimental design facilitated the examination of female, male and female–male interacting effects on sperm behaviour in follicular fluid. Each experimental block comprised the follicular fluid and sperm samples from a unique set of two couples, exposing sperm from each male to follicular fluid from their partner and a non-partner (figure 1a,b). We performed two cross-classified experiments that differed in how sperm experienced the choice of follicular fluid from different females, being either ‘simultaneous' or ‘non-simultaneous' (figure 1c,d). Sperm responsiveness to follicular fluid was quantified by counting the number of sperm accumulating in the follicular fluid from each female. Each assay was replicated twice and repeatability among experimental replicates was high (simultaneous choice experiment: R = 0.96, 95% CI = 0.92–0.98, p < 0.001; non-simultaneous choice experiment: R = 0.97, 95% CI = 0.95–0.98, p < 0.001, see electronic supplementary material). High repeatability is consistent with a chemotactic response rather than mechanical trapping effects, where sperm accumulate due to adsorption to the experimental apparatus, which is unlikely to exhibit high repeatability between experimental replicates [21]. Figure 1. Microscopic mate choice. An overview of the experimental design used to assess variation in sperm accumulation responses to follicular fluid from different females. (a) Each experimental block consisted of samples of follicular fluid and sperm that were obtained from two couples—couple 1, comprising female 1 (F1) and male 1 (M1), and couple 2, comprising female 2 (F2) and male 2(M2)—undergoing clinical assisted reproductive treatment. (b) Sperm accumulation in follicular fluid was assessed by crossing females and males in all possible combinations for each experimental block in the cross-classified design. Each cross was replicated twice for every female–male combination. Thus, in each experimental block, sperm were exposed to follicular fluid from both their partner (from the same couple) or a non-partner (from a different couple). An example of a single block in the experimental design is presented, where sperm were exposure to follicular fluid (housed in microcapillary tubes) from two females under (c) simultaneous (n = 8 blocks) or (d) non-simultaneous (n = 22 blocks) experimental conditions. The microcapillary tubes were sealed with a plug (indicated in grey) at the end of the tube where sperm were added to the Petri dish. Thus, to enter the microcapillary tube, sperm had to swim the length of the microcapillary tube. In the simultaneous experimental design sperm were presented with a simultaneous choice of follicular fluid from two females, while in the non-simultaneous experimental design sperm were presented with a choice of follicular fluid from one female (either the partner or non-partner) and a control medium. The number of sperm that successfully entered the microcapillary tube was counted by light microscopy at 300× magnification to quantify sperm accumulation in and responsiveness to the follicular fluid. (Online version in colour.)
The simultaneous choice experiment consisted of 16 couples, comprising eight blocks of factorial crosses (14 IVF and 2 ICSI treatments; note that excluding ICSI patients from the analyses did not qualitatively alter our findings; see below). In each block, sperm were presented simultaneously with follicular fluid in 2 μl microcapillary tubes in a Petri dish from two females (a partner and a non-partner) to determine if sperm preferentially and consistently swim towards and accumulate in the follicular fluid of a specific female (figure 1c; electronic supplementary material). Thus, sperm had to swim the length of the microcapillary tube (approx. 30 mm), moving up the chemoattractant gradient, to enter the microcapillary tube containing the follicular fluid. Following sperm addition, Petri dishes were left undisturbed in the incubator for 1 h to allow the sperm time to migrate towards and accumulate within the microcapillary tubes. We evaluated sperm responses to follicular fluid when exposure to follicular fluid was non-simultaneous (figure 1d): sperm were presented with the choice between the follicular fluid from one female (either the partner or non-partner) and a control sperm preparation medium (SpermRinse) solution (n = 44 couples, 22 blocks of factorial crosses; 30 IVF and 14 ICSI treatments). This experimental design allowed us to test if sperm preferentially and consistently accumulate in the follicular fluid of a specific female relative to a control solution. Assays were performed in Petri dishes primarily as described in the ‘Simultaneous choice of follicular fluid experiment’ section (see electronic supplementary material for details of the minor modifications in the experimental design), although in the case of the non-simultaneous choice experiment only one female's follicular fluid was presented in each Petri dish (figure 1d). One hour after sperm addition, the microcapillary tubes were removed from the Petri dish and placed on a microscope slide (Menzel-Glaser, Braunschweig, Germany). The number of sperm present in the microcapillary tubes were counted under 300× magnification using an Olympus IMT2 inverted microscope. Because microcapillary tubes are three-dimensional objects, we adjusted the plane of focus as the field of view moved along the length of the microcapillary tube to ensure that sperm were counted on all focal planes of the tube. Variance in sperm responses to follicular fluid can be caused by differential sperm chemotactic responses and/or differential responses in sperm swimming speed (i.e. chemokinesis). To evaluate the potential role of chemokinesis in mediating sperm responses to follicular fluid we characterized sperm performance in 12 males from a subset of six experimental blocks from the non-simultaneous choice experiment. Sperm swimming characteristics were quantified using computer-assisted sperm analyses (CASA). Sperm swimming characteristics from each male were assessed under three different treatments: in follicular fluid from their partner, follicular fluid from a non-partner, and in sperm preparation medium (SpermRinse) as a control (see electronic supplementary material). To investigate whether sperm respond differentially to follicular fluid we performed a series of sequential two-way analysis of variance (ANOVA) models to estimate the female, male and female–male interaction effects on sperm accumulation in the simultaneous (n = 8 blocks) and non-simultaneous (n = 22 blocks) choice experiments. Analyses were performed separately for each experimental block of factorial crosses and then combined in Microsoft Excel (v.16.16.13) into a final model using a ‘North Carolina II' block design [20]. In the simultaneous choice experiment, we treated sperm accumulation (i.e. the number of sperm counted) in the microcapillary tube housing the follicular fluid as the response variable. In the non-simultaneous choice experiment, sperm accumulation was almost 10 times greater in follicular fluid (435.0 ± 39.0) than the control solution (45.1 ± 3.0, generalized linear mixed model (GLMM): fixed intercept: Z = 12.2, p < 0.001), confirming the chemoattractant properties of follicular fluid. However, as our aim was to assess how sperm responds to follicular fluid of different females, we treated sperm responsiveness to follicular fluid, quantified as the difference in the number of sperm accumulating in the microcapillary tube containing the follicular fluid relative to the microcapillary tube containing the control solution, as the response variable in the non-simultaneous choice experiment. In both the simultaneous and non-simultaneous experiments, we avoided interpreting significant female or male main effects in the presence of a significant female–male interaction. When blocks containing ICSI patients were removed from the analyses we obtained qualitatively similar results (see electronic supplementary material, table S1) and therefore we present results from the full dataset. We next examined if sperm accumulation and responsiveness are influenced by the origin of the follicular fluid (i.e. partner versus non-partner follicular fluid). In the simultaneous choice experiment, we fitted a GLMM with a logit link function, treating sperm accumulation (i.e. the number of sperm counted) in either the partner or non-partner follicular fluid, which represented a binary choice of simultaneously presented follicular fluid (figure 1c), as a binomial response variable. The simultaneous choice model was fitted with fertility treatment (IVF versus ICSI) as fixed effects and the female, male, and female–male interaction, the experimental block and observation number (to account for overdispersion) as random effects. In the non-simultaneous choice experiment, we fitted a LMM treating sperm responsiveness (log10 transformation on positivized values) as the response variable, with follicular fluid origin (partner versus non-partner) and fertility treatment (IVF versus ICSI) as fixed effects (note the non-significant interaction term between the categorical fixed effects was dropped from the model and sperm density was removed from the final model as inclusion of this variable impaired model fit), and female, male, female–male interactions, and experimental block as random effects. Note that we obtained qualitatively similar results when we assessed if sperm accumulation and responsiveness were influenced by the origin of follicular fluid using simplified models where the number of random effects present in our main analyses were reduced (see electronic supplementary material). We then explored if sperm swimming behaviour was influenced by follicular fluid (compared to a control solution) and if sperm behaviour differs when swimming in follicular fluid from a partner compared to a non-partner. We used principal components (PC) analysis as a data reduction method to reduce the seven highly correlated sperm swimming parameters produced by CASA into two PC's with eigenvalues greater than one (electronic supplementary material, table S2). To assess if sperm swimming speed differs when swimming in follicular fluid compared to a control solution, we fitted a LMM with experimental medium (follicular fluid versus a control solution), fertility treatment (IVF versus ICSI), and their interaction as fixed effects and male identity and experimental block as random effects. Next, we used a LMM to examine if sperm swimming speed is influenced by the origin of the follicular fluid (partner versus non-partner), treating follicular fluid origin, fertility treatment (IVF versus ICSI) and their interaction as fixed effects, while including male identity and experimental block as random effects. To assess if sperm responsiveness to follicular fluid was influenced by variance in sperm swimming speed among males, we fitted a LMM with sperm responsiveness (sperm in follicular fluid—sperm in control) as the response variable, with sperm swimming speed (PC1 and PC2), the density of sperm added to the Petri dish (which is positively related with sperm accumulation in the microcapillary tubes, LMM: χ2 = 15.2, p < 0.001) and fertility treatment (IVF versus ICSI) as fixed predictor variables and male identity and experimental block as random effects. Finally, we examined whether follicular fluid that preferentially attract sperm from their partner (relative to a non-partner) had higher fertilization/embryo quality/pregnancy/live birth success during IVF treatment than follicular fluid less capable of attracting sperm from their partner. All sperm data were derived from replicate mean values (see electronic supplementary material). We treated the proportion of eggs fertilized, pregnancies success (0 or 1) and live birth success (0 or 1) as binomial response variables and fitted GLMMs with a logit function. In the simultaneous choice experiment, partner sperm preference (i.e. the difference in sperm accumulation to the partner versus non-partner follicular fluid) was treated as a continuous fixed effect, and the experimental block and observation number included as random effects. In the non-simultaneous choice experiment, partner sperm responsiveness (i.e. partner follicular fluid—control sperm count) was treated as a continuous fixed effect, and the experimental block and observation number included as random effects. For embryo quality, we used LMMs with partner sperm preference or partner sperm responsiveness as predictor variables for the simultaneous and non-simultaneous choice experiments, respectively, and fitted models with experimental block as a random effect. Analyses of these fitness variables excluded patients treated using ICSI, as fertilizations using ICSI involve sperm being injected into the egg rather than sperm-directed movement towards the egg. All analyses were performed in R Studio v.1.1.463 [22], with LMM and GLMM models fitted using the lme4 package [23]. In GLMM and LMM models, parameters were estimated using the Laplace approximation of log-likelihoods and the Satterthwaite's method, respectively. GLMMs were initially fit with the bobyqa optimizer, but in cases where model convergence failed we set the nAGQ scalar to zero. Model diagnostics were performed by assessing overdispersion using the RVAideMemoire package in R [24] and testing for uniform distribution of the scaled residuals in the DHARMa package in R [25]. When sperm were presented with a simultaneous choice of swimming towards follicular fluids from two females (a partner and a non-partner, n = 16 couples, eight blocks of factorial crosses; figure 1c), sperm accumulation in follicular fluid was significantly influenced by the interactive effect of female–male identity (F8,32 = 19.38, p < 0.001; figure 2a, table 1a). However, in internally fertilizing species such as humans, sperm are never presented with the simultaneous choice of follicular fluid from more than one female. Therefore, we performed a second cross-classified experiment under biologically relevant conditions, where sperm were non-simultaneously exposed to follicular fluid from two females (n = 44 couples, 22 blocks of factorial crosses; figure 1d). In the non-simultaneous choice experiment, sperm were given the choice between the follicular fluid from one female (either the partner or non-partner) and a control solution (sperm preparation medium). When sperm were presented with the non-simultaneous choice of follicular fluid, sperm responsiveness was also influenced by the interaction between female and male identity (F22,88 = 21.82, p < 0.001; figure 2b, table 1b). The significant interactive effects of female–male identity on sperm behaviour remained when we examined IVF and ICSI patients separately in the simultaneous and non-simultaneous choice experiments (electronic supplementary material, table S1). Figure 2. The effect of female, male and female–male interactive effects on sperm accumulation in the (a) simultaneous and (b) non-simultaneous choice experiments. The effect size (Cohen's d) and 95% confidence intervals are presented for each effect in the cross-classified design. Plots are for illustrative purposes only, as values were derived from standard F-test effect size calculations. Effects are considered significant when confidence intervals do not overlap with zero.
To evaluate the potential role of partner effects, we assessed if the female–male interactive effects in sperm accumulation/responsiveness are influence by the origin of follicular fluid. In the simultaneous choice experiment, sperm accumulation did not differ between follicular fluid of the partner (515.7 ± 79.9, mean ± s.e.) or non-partner (441.9 ± 51.0, GLMM; fixed intercept: Z = −0.29, p = 0.77; fertility treatment effect (IVF versus ICSI): Z = 0.71, p = 0.48), indicating that sperm do not preferentially accumulate in the follicular fluid of their partner. Similarly, when we assessed sperm responsiveness to follicular fluid in the non-simultaneous choice experiment, patterns of sperm responsiveness were not affected by the origin of the follicular fluid (follicular fluid origin: χ2 = 0.32, p = 0.57; fertility treatment: χ2 = 3.23, p = 0.07). Although follicular fluid influenced sperm behaviour in the patients we considered (electronic supplementary material, table S3), sperm swimming behaviour did not differ when exposed to follicular fluid from either a male's partner or a non-partner (electronic supplementary material, table S3). Moreover, sperm responsiveness to follicular fluid was not related to sperm swimming speed, suggesting that patterns of sperm accumulation are not explained by sperm chemokinetic responses (sperm velocity PC1: χ2 = 0.25, p = 0.61; sperm velocity PC2: χ2 = 0.01, p = 0.94; fertility treatment: χ2 = 0.03, p = 0.87; sperm density: χ2 = 1.42, p = 0.23). We found limited evidence that fitness measures were influenced by sperm responses to follicular fluids. In the simultaneous choice experiment, fertilization rates using IVF were higher when sperm were more responsive to their partner's follicular fluid (i.e. partner sperm preference was stronger, Z = 2.25, p = 0.02, electronic supplementary material, figure S1a). Similarly, there was a statistical trend suggesting that embryo quality increased when partner sperm preference was stronger (χ2 = 3.51, p = 0.06). However, these relationships were driven entirely by two outlying data points (see electronic supplemental material, figure S1a). Partner sperm preference did not predict whether IVF treatment resulted in clinical pregnancy (Z = 0.82, p = 0.40) or live births (Z = 1.11, p = 0.27). In the non-simultaneous choice experiment, partner sperm responsiveness did not predict the proportion of eggs fertilized using IVF (Z = 0.52, p = 0.60, electronic supplementary material, figure S1b), embryo quality (χ2 = 1.28, p = 0.26), clinical pregnancy (Z = −0.80, p = 0.42) or live birth success (Z = −1.34, p = 0.18). Chemical communication between eggs and sperm is critical for fertilization. As sperm make their way towards eggs, sperm behaviour is influenced by signals released from unfertilized eggs and/or the female's reproductive tract, leading to the traditional view that chemical signals act only to guide sperm to eggs in internal fertilizers, like humans [5]. Our findings challenge this long-standing paradigm. We found concordant patterns of sperm responses to follicular fluid under two distinct experimental conditions, performed with two independent groups of couples that were temporally separated. Our results demonstrate that sperm accumulation in follicular fluid depends on the specific combinations of follicular fluid and sperm, and that follicular fluid preferentially attracts sperm from specific males. The non-random, repeatable sperm accumulation responses we detected suggest that chemical communication between eggs and sperm allows females to exert ‘cryptic choice' over which sperm fertilize their eggs. Female–male interactive effects during reproduction are a hallmark of cryptic female choice [10]. Such differential sperm responses to follicular fluid have the potential to influence fertilization success between specific female–male partners. Sperm number is a key determinant of fertilization success [26]. In humans, a minute fraction of ejaculated sperm makes its way up the fallopian tube to the site of fertilization (mean = ∼250 sperm [27]). Of these few remaining sperm, roughly one in ten is capacitated (a biochemical process required for fertilization capacity) and capable of responding to chemoattractants and fertilizing the egg [28]. The ever-dwindling number of sperm capable of fertilizing eggs as they move through the female's reproductive tract suggest that the capacity for chemoattractants to differentially recruit sperm from specific males could make the difference in ensuring fertilization success. Yet, despite the potential for differential chemotactic responses to influence fertility, we found only weak evidence that sperm responses to chemoattractants influence fertilization success and later fitness measures. This contrasts with findings in marine invertebrates, where differential sperm responses to chemoattractants can influence both fertilization success and subsequent embryo viability [6,29]. However, the lack of a clear relationship between partner sperm preference and/or partner sperm responsiveness and fitness measures during IVF is perhaps not surprising given the constraints of the clinical setting where our experiments took place. Clinical practices place sperm in close contact with eggs, potentially minimizing the importance of chemoattractants prior to fertilization, and attempt to maximize fertilization success by using sperm concentrations several orders of magnitude greater than are found in the fallopian tube at the site of fertilization in vivo (e.g. typically 20 000 sperm per egg, [17]). Downstream clinical treatment of embryos following fertilizations also removes potentially important biological processes. Thus, while challenging to detect in vitro, follicular fluid-mediated differential recruitment of sperm could play an important role during in vivo fertilizations in humans, although this requires further validation. An important next step is to determine if incorporating considerations of chemical communication between gametes into clinical practices could improve not only fertilization success but the quality of developing embryos both prior to and post-implantation. Female–male interactive effects are also characteristic of mate choice for genetic compatibility generally [3,4,6,7]. Thus, the female–male interactive effects we detected raise the possibility that preferential sperm accumulation reflects a chemoattractant-mediated mechanism occurring prior to direct sperm-egg interactions to avoid post-mating genomic incompatibilities (sensu 12). For example, sperm could swim preferentially towards the follicular fluid from their partner, provided human pairing reflect mate choice for genetic compatibility at the major histocompatibility complex (MHC), a diverse chromosomal region that functions in immune defence (although whether this is the case remains controversial, [30–34]). Under this scenario, sperm responses to their partner's (or indeed non-partner's) follicular fluid may reflect the degree of genetic compatibility between the pair. Alternatively, as the couples in our study were undergoing fertility treatment, the interactive effect between males and females could stem from a clinical pathology that impairs chemical communication between gametes. This has direct clinical relevance as a high proportion (32%) of couples undergoing fertility treatment in the UK have a cryptic (‘unexplained' or idiopathic) cause to their infertility [35]. In cases of idiopathic infertility, sperm may swim preferentially towards follicular fluid from random (i.e. non-partner) females over follicular fluid from their partner. However, we find no support for either of these possibilities as female–male interactive effects were not explained by differential responses to either partners or non-partners. Nevertheless, considering female–male interactive effects when examining the mechanistic underpinnings of chemical communication between gametes will help to clarify the factor(s) influencing sperm accumulation in follicular fluid. Our results demonstrate that patterns of sperm accumulation are shaped by combinatory female–male effects and cannot be explained by differences in the quality and/or amount of chemoattractants present in the follicular fluid of different females or by differences in male ejaculate quality. Female–male interactive effects are a key diagnostic required for demonstrating cryptic female choice of sperm [12]. Thus, despite ample scope for humans to exercise pre-mating mate choice [2,30], chemosensory communication between gametes retains a role in selectively recruiting sperm. Indeed, in their initial demonstration that human sperm are attracted by chemical signals in follicular fluid more than 25 years ago, Ralt et al. [16] reported variation in sperm responsiveness to follicular fluids from different females. However, until now, the implications of this finding have not been explored. Our results imply that chemoattractants may allow females to exert post-mating (i.e. cryptic female choice) gamete-mediated mate choice. These findings extend the traditional view that chemoattraction solely plays a role in increasing sperm-egg interactions in humans [5] and instead suggest that chemical communication between eggs and sperm may also have a sexually selected role. A critical next step is to determine if such female–male interactive effects are a common feature of mammalian reproduction, including humans not undergoing assisted fertility treatments (although this is logistically and ethically challenging), and to examine the potential for gamete-mediated mate choice to influence embryo quality under biologically relevant conditions. Nevertheless, our findings suggest that chemosensory-driven interactive responses to chemoattractants probably span the animal tree of life and potentially provide a widespread mechanism of gamete-mediated mate choice that is currently underappreciated. Clarifying the significance of chemical communication between human gametes during fertilizations and uncovering the molecular mechanisms influencing differential sperm response may aid in the development of new approaches for diagnosing and treating unexplained infertility and improving the efficiency and safety of assisted reproductive treatments. The dataset associated with this study is available from the Dryad Digital Repository: https://doi.org/10.5061/dryad.wdbrv15kb [36]. J.L.F. and D.R.B conceived the study, obtained funding and drafted the manuscript. C.W., A.D. and A.Y. performed the majority of the technical work in collecting the data with supervision from J.L.F., H.R.H. and M.C. J.L.F., C.W. and A.D. analysed the data. All authors helped to draft the manuscript and read and approved the final manuscript. We declare we have no competing interests. Research was supported by the Manchester University NHS Foundation Trust, the University of Manchester, the National Institutes of Health Research, a Knut and Alice Wallenberg Academy Fellowship (2016-0146) and Swedish Research Council Grant (2017-04680) to J.L.F., and a Wenner Gren Postdoctoral Fellowship to A.D. We thank the patients who consented to their sperm and follicular fluid being used in this research and the staff at the Department of Reproductive Medicine, St Mary's Hospital, Manchester for making this study possible, particularly Claudette Wright and Chelsea Buck. We also thank Andrea Pilastro for use of his sperm tracker. FootnotesElectronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.5004644. © 2020 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original author and source are credited. References
Page 15Human activity often results in rapid and extreme environmental variation with biologists becoming increasingly interested in predicting how populations will respond to such novel changes [1]. The capacity for developmental changes in the expression of adaptive phenotypes (i.e. evolvability) could be key for allowing animals to persist in anthropogenic environments [2]. For example, cities may represent extremely novel environments with alterations of food availability, predation, spatial structure, lighting period, community structure, as well as auditory and visual cues. While there is ample evidence of phenotypic divergence between rural and urban populations of animals, we still know little about how selection could be differentially operating between these environments across a suite of functionally salient traits, and across replicate habitat gradients in other cities [3]. Thus, while phenotypic divergence occurs between urban and rural habitats, it could be that some changes are due to variation in cognitive demands, while others are due to variation in biomechanical demands. Indeed, prior research suggests that a larger brain size can facilitate the innovative behaviours that could allow animals to persist in urban habitats, including the exploitation of novel foods, or the evasion of novel threats [1–3, but see 4]. Characterizing the morphological changes during a transition from a rural to urban environment could indicate how wild populations specifically cope with living in close proximity to humans but may also provide insights into how domestication is initiated. Domestication results in a number of stereotypical changes known as ‘domestication syndrome’ [5–9]. Domestication syndrome refers to the suite of phenotypic changes that are known to occur in response to the domestication process. For example, domestication leads to stereotypical changes across species toward more docile behaviour, coat colour changes, reduced total brain size, reductions in tooth size, prolongations of juvenile behaviour, and changes in craniofacial traits, including a shortened skull morphology. Such changes are well documented among several taxa, with foxes and domestic dogs being particularly well studied [9–12]. Domestic dogs underwent extensive morphological change to form today's modern dog breeds with especially notable changes in the skull that are primarily characterized by snout lengthening and shortening [12]. While it should be noted that breed formation is distinct from, and a secondary outcome that would follow from domestication, these changes across domestic dog breeds could represent the magnification of an initial trend. Indeed, while such differences are often solely attributed to artificial selection, it is also likely that developmental biases are present [13] for canid skulls that direct variation toward such changes. In fact, the dog skull (both domesticated and wild), and that of other canids is known to be modular with the anterior snout region forming a separate variational module from the brain case [12,14]. Such modularity makes it possible for changes in the snout to be independent of the rest of the skull. If these patterns of modularity are deeply ancestral in canids, they could emerge as phylogenetic effects that bias the evolution of other species at a macroevolutionary scale, but also affect change at a microevolutionary scale by limiting the number of possible phenotypes. Indeed, evidence from the well-known Belyayev domestication experiments in red foxes, and which solely favoured behavioural ‘tameness’, was intended to mimic the selection regime during the initial domestication process of dogs but has also resulted in a number of morphological changes including a relative shortening of the snout [9,10,15]. Regarding domestic dogs, the conditions during the initiation of their domestication are largely unknown. One hypothesis suggests that behavioural changes were a driver of initial changes toward domestication. Indeed, grey wolves (Canis lupis) are social pack animals and, similar to dogs, are especially noted for their ability to convey and interpret facial expressions. Evidence suggests that wolves that were in early contact with humans developed shorter, wider skulls thought to be more interpretable by humans [5,16,17]. Alternatively, such morphological changes could simply be present due to functional demands caused by changes in diet that correspond with the presence of humans. ‘Scavenging’ partly processed carcasses or cooked food from humans could have reduced the stresses in wolves' jaws, thus effecting morphological change. Therefore, investigating a canid that is less social (solitary hunters but monogamous and sometimes living in small family groups) with populations showing very recent close proximity to, but with few social interactions with humans, could be particularly informative for discerning the initial primary drivers of skull shape divergence (i.e. a surrogate of the conditions that possibly initiated dog domestication). To address these issues, we focused on testing for morphological divergence in the skull of red foxes (Vulpes vulpes) inhabiting rural and urban habitats in southern England. Starting over a century ago urban foxes have been recorded in many British cities, such as Birmingham, Bristol and London [18]. Urban foxes appear to have made a significant ecological shift as they now exploit shelter and can have upwards of 37% of their diet consisting of scavenged food [19]. In turn, urban foxes show substantially reduced home ranges in urban habitats relative to rural ones (0.4 km2 versus 30 km2 for urban and rural habitats, respectively), suggesting barriers to gene flow could exist and provide an opportunity to adapt to local conditions [19,20]. Indeed, previous research has suggested that urban foxes in Switzerland are somewhat genetically isolated from their rural conspecifics [21,22]. Potential morphological differences between urban and rural foxes are currently unknown but their skull provides a complex multivariate trait that could provide insights into functional differences and evolutionary mechanisms of differentiation. Using the V. vulpes skull, we tested the general hypotheses that differential conditions (selective pressures or plasticity) between urban and rural environments would produce changes in skull morphology that reflect differences in ecology. Specifically, in line with trends found within domestication syndrome, we predicted that urban environments would favour a skull with a shorter wider snout. Additionally, in line with previous findings from mammals (but against the predictions of domestication syndrome), we predicted that urban environments would be associated with a larger brain (and hence larger brain case) due to increased cognitive demands [1]. Also, in line with other examples of habitat divergence between urban and rural environments, we hypothesized that sex differences in the divergence of skull morphology would arise between environments given differences in life history demands [23–25]. If present, we predicted that sexual dimorphism would be reduced in the urban environment in line with domestication syndrome [9]. Finally, as an alternative driver of divergence patterns we accounted for possible phylogenetic effects and developmental biases that could influence outcomes across an urban/rural habitat gradient. Specifically, we tested whether patterns of divergence between species of Vulpini were influenced by phylogeny. We also then tested whether patterns of divergence were aligned among micro- and macroevolutionary scales. Understanding these aspects of divergence in response to anthropogenic factors could greatly increase our ability to predict the responses of other animal populations to human environments, while also informing hypotheses surrounding the initiation and outcomes of domestication. A total of 111 skulls of red foxes (Vulpes vulpes) were available from London (n = 75, 38 females, 37 males) and the surrounding boroughs (n = 36, 17 males, 19 females), which are housed in the collections of National Museums Scotland (NMS); the location of collection, and sex of each individual are provided in electronic supplementary material, table S1. These were collected from 1971 to 1973 by Steve Harris [26]. All specimens have information about date and collecting locality, which allowed for classification of individuals into urban and rural locations although no information about relatedness was available. Locations were checked against contemporary OS maps to determine whether they were rural or urban at that time, because many locations have become urbanised since the time of collection. Urban collection sites were classified as those containing buildings, street lighting and lacking wooded areas (notably collection sites most often include a precise street name). Rural sites were dominated by wooded areas and little to no human development at time of collection. While more refined methods are available for classification of habitats [27], this was not possible with our data due to a lack of information at the time of collection. Nonetheless, our approach was in line with previous studies assessing rural–urban differences in mammals [28]. As growth in red foxes is normally complete after a year, only adult specimens, with fused basi-sphenoid sutures, were kept for analysis to limit variation due growth allometry [29]. To quantify the morphology of fox skulls, digital photographs were taken of each specimen using a Nikon Coolpix 4500 (Nikon, Japan). Specimens were placed on modelling clay to standardize their articulation for photography. For each specimen, an image of the dorsal and ventral aspects of each skull was collected for landmarking. Briefly, the dorsal aspect of the skull was used to define 36 homologous landmarks, while the ventral aspect was used to define 29 landmarks and followed a similar protocol to Drake & Klingenberg [12] (figure 1; electronic supplementary material, table S2). Figure 1. Landmarks for dorsal (a) and ventral (b) aspects of a red fox, Vulpes vulpes, skull. Photographs: Neil McLean (copyright National Museums Scotland).
Across the Vulpini clade, we collected similar data from a further 163 specimens housed both at the National Museums Scotland and the Natural History Museum (London). This dataset consisted of 12 species with 10 from Vulpes (V. cana, V. chama, V. corsac, V. ferrilata, V. macrotis, V. lagopus, V. rueppellii, V. velox and V. zerda), and two basal species (Otocyon megalotis and Nyctereutes procyonoides) with varying sample sizes and used the same landmarking protocol as above (electronic supplementary material, table S3). Relationships among Vulpini and the additional canids were based on a recently published supertree [30]. Landmark data for each of the ventral and dorsal aspects of each skull were corrected for variation in size and orientation using a generalized Procrustes analysis that included all specimens for each aspect. Partial warp scores, which accounted for quantitative variation in shape, were collected for each of the ventral and dorsal aspects for further statistical analysis. The steps in this analysis involved the use of tpsUtil to append all specimens into a single tps file, and tpsRelw to perform Procrustes transformation, thin-plate spline projection, and the extraction of partial warp scores [31]. Partial warp scores are amenable to multivariate statistical analyses and represent the rotation of Procrustes residuals around the Procrustes mean configuration [32]. To assess the influences of sex and habitat class on skull shape a MANOVA was performed for each skull aspect of the red fox data using base functions in R v. 3.5.0 [33]. Sex and habitat (urban/rural) and their interaction were used as explanatory variables, while partial warp scores were the shape response variables. To determine whether the degree of sexual dimorphism differed between habitats, we compared the magnitudes of shape difference between males and females from rural and urban sites. This analysis relied on measurements of sex-based Procrustes distances defined as the square root of the sum of squared differences in the positions of the landmarks in two shapes [31]. The Procrustes distance between male and female foxes from the rural habitat was compared to the corresponding male/female Procrustes distance derived from the urban habitat using 900 bootstraps [31, p. 224]. This produced 95% confidence intervals for both urban and rural groups, with a lack of overlap indicating a significant difference. This analysis was performed using the Coordgen8 package to format landmark data files in conjunction with TwoGroup8 to perform the bootstrapping procedure [31,34]. To visualize shape variation for biological interpretation, we produced a series of deformation grids depicting the two-dimensional effects of sex and habitat on skull shape for red foxes. These deformation grids were based on canonical variables derived from our MANOVA models using the candisc package in R v. 3.5.0. [35]. Specifically, while canonical variables are traditionally limited to a one-way MANOVA design, this package allowed for the generalization of our two-way MANOVA designs. Therefore, our canonical scores for each of sex and habitat take account of the other factor in the model. Deformation grids for sex and habitat effects were created using these canonical scores as an independent variable in a multivariate regression on coordinate data using tpsRelw [36]. We assessed whether phylogenetic effects influenced fox skull shape. Upon image collection it was apparent that a wide degree of size variation was present among species of foxes. Therefore, to quantify and account for allometry in Vulpini skull shape, we performed a Procrustes ANOVA between centroid size and skull shape. We found a significant effect of allometry (dorsal aspect r2 = 0.11, F = 20.48, p = 0.001; ventral aspect: r2 = 0.12, F = 22.11, p = 0.001). Thus, we then performed a regression of shape on geometric centroid size for both dorsal and ventral aspects of the skulls to generate an allometry-minimized landmark dataset based on residuals [29]. Finally, to minimize effects from differential sample sizes across species, we calculated mean landmark configurations for each Vulpini species, the two basal species, and then performed a principal component analysis (PCA, see electronic supplementary material, table S4) on both the individual and mean Procrustes-transformed shape data to quantify variation among skulls. We used the geomorph package (v. 3.0.1) in R to conduct tests for allometry, and to perform PCA analysis following general Procrustes superimposition [37]. All phylogenetic comparative methods were performed using a time-calibrated, species-level supertree of the Carnivora [30]. We extracted the relationships of the 12 Vulpini species from this supertree, for which we had dorsal and ventral morphometric data, and pruned all remaining taxa using the ape package in R [38]. To assess whether divergence across the urban/rural habitat axis was similar to trends found across the phylogeny we performed a series of steps. First, we used the canonical axis of habitat divergence generated above (in candisc) using each of the ventral and dorsal aspects (figure 1). This canonical axis and its scores represented microevolution across the urban/rural habitats while taking account of sexual dimorphism. Second, to determine the major axis of variation for Vulpini (i.e. macroevolution), we extracted the first principal component from the mean shape data of the Vulpini clade. Third, we used a multivariate version of Blomberg's K to estimate the degree of phylogenetic signal across the Vulpini clade in our PC score data (axes 1–6) from the dorsal and ventral aspects [39]. Blomberg's K measures phylogenetic signal by quantifying the amount of observed variance in dorsal and ventral PC scores relative to variance expected under Brownian motion. K ranges from 0, whereby no phylogenetic signal is detected and closely related taxa exhibit traits that, on average, are not more similar than more distantly related taxa, to infinity. When K = 1 the trait exhibits strong phylogenetic signal and is evolving under a model of Brownian motion. When K > 1 closely related taxa exhibit trait values more similar than would be expected under Brownian motion [39]. We tested whether K significantly differed from 0 (i.e. a sign of no phylogenetic signal) by comparing our value of K to a null distribution of K values generated via 1000 simulations on a star phylogeny, which serves to remove or eliminate phylogenetic signal by rescaling branch lengths [39]. We used the K.mult function in the R package phylocurve (v. 2.0.9) to conduct our multivariate assessment of Blomberg's K [40]. Performing evolutionary analyses on a dataset with a small number of species can result in greater error rates depending on data structure [41–43]. To determine the degree of statistical power in our analysis of K, we report the value of estimated power between the simulated data and our own given by the K.mult function [40]. It was qualitatively apparent that the magnitude of divergence was several fold greater among species across Vulpini than between the urban/rural habitat axis. However, both types of divergence (micro- and macroevolution) could follow a common trajectory. Therefore, we compared the major microevolutionary trajectory of skull shape divergence with the major trajectory of macroevolutionary divergence. Quantitatively, this involved extracting the main trajectories for each type of divergence (macro- and micro-). For divergence between urban and rural habitats, the canonical scores, calculated using habitat as a grouping variable (and controlling for sex variation), were used to provide a microevolutionary trajectory. To represent divergence among Vulpini, we used the major axis of divergence (i.e. PC1) derived from the landmark dataset comprised of each species' mean shape. Specifically, because of clear differences in magnitude comparing trajectories of micro- and macroevolutionary divergence required a scale-free approach. To derive a scale-free vector of microevolutionary divergence for comparison in shape space, we regressed the Procrustes superimposed landmark data from the dorsal and ventral aspects of urban/rural populations on their habitat-derived canonical axis [31, p. 257]. Similarly, the vector of macroevolutionary divergence was calculated by regressing Vulpini landmark data against PC1. The scale-free observed angle between these vectors for micro- and macroevolutionary divergence (for both dorsal and ventral aspects) was then calculated as the arc cosine. We then ran 900 bootstraps with replacement for each group (urban/rural and Vulpini) independently to produce 95% confidence intervals. The observed angle between micro- and macroevolutionary divergence was compared against the confidence interval of angles to determine whether it differed from random processes (i.e. did the observed angle lie outside the confidence interval?). These procedures were performed using standard routines within the software Regress8 [34]. We also performed complementary approaches through an alternative procedure based upon linear model evaluation with a randomized residual permutation procedure [44]. While allowing for similar comparisons of trajectory this approach also allowed for tests of differences in the magnitude of evolutionary divergence along a common trajectory. Specifically, this was performed using our landmark data with habitat and vectors derived from our PC1 and DFA scores as explanatory variables and using 1000 permutations using the pairwise function within the RRPP package in R [44]. We found strong evidence that skull shape was different between urban and rural habitats. For both the ventral and dorsal aspects, habitat had a large effect on shape (table 1 and figure 2). Habitat also interacted with sex in both views, although with a strong but slightly smaller effect on shape relative to habitat. Furthermore, sex alone had a major effect on shape, especially within the dorsal aspect, where its impact was larger than habitat (table 1). The degree of sexual dimorphism also differed between habitats, with the dorsal aspect showing a significant 28% reduction in dimorphism in the urban habitat (95% CIs did not cross zero). Anatomically we found widespread differences in skull shape between habitats, with urban foxes having a noticeably shortened wider snout with a reduced maxillary region relative to rural foxes (figure 3). However, the tip of the snout, which is comprised of the premaxillary and nasal regions, showed some degree of widening in urban foxes, which was especially evident from the ventral view. Finally, the sagittal crest was extended posteriorly in urban foxes, while the zygomatic region was relatively reduced in terms of both length and width, along with the braincase. Many of these shape changes could be related to the development of jaw muscles [45,46]. While an extended sagittal crest would indicate an increased area of attachment of the temporalis muscle and indicate a higher bite force, a gracile zygomatic arch would also indicate a reduced masseter muscle in urban foxes. Indeed, finite-element modelling of biting in canids has demonstrated that the zygomatic region experiences particularly high stresses [46], and so reinforcement in this region may be adaptive. Figure 2. Frequency histograms depicting the statistical discrimination of habitat and sex-based differences in dorsal and ventral red fox skull shape. In (a,b), the habitat-based differences are depicted for the dorsal and ventral views, respectively, (light grey = rural, black = urban). In (c,d) the sex differences from the dorsal and ventral views are shown respectively (light grey = female, black = male).
Figure 3. Skull shape variation in red foxes (Vulpes vulpes) in relation to urban and rural habitats from the dorsal and ventral aspects. Trends are magnified by 3× to enhance the interpretation of shape variation. Note the snout (LMs 1–14 in the dorsal aspect, 1–10 in the ventral view) containing the maxillary region and nasal regions, and the braincase (LMs 25–36 in the dorsal aspect, 16–29 in the ventral aspect) containing the sagittal crest.
Such anatomical variation is likely to provide a number of ecologically functional differences between urban and rural populations. Firstly, a shorter snout, as found in urban foxes, should confer a higher mechanical advantage but with reduced closing speed of the jaw [46]. This may be advantageous in an urban habitat where resources are more likely to be accessed as stationary patches of discarded human foods. Furthermore, in some cases, these foods may require a greater force to access them, explaining the expanded sagittal crest in skulls of urban foxes. Consistent with this the squamous temporalis is expanded in urban foxes as indicated from the ventral aspect (figure 3). In a rural habitat, an increased jaw-closing speed would be conferred by an increase in its length and aid in capture of motile prey, e.g. voles, mice and rabbits. While having an overall smaller snout the increased nasal region in urban foxes (figure 3, at the tip of the snout) may also reflect their ecology, which could be more dependent on olfactory cues than other senses. Contrary to our prediction the braincase appeared to be smaller in the urban habitat. While this might suggest a smaller brain (in agreement with domestication syndrome), it could possibly reflect changes in biomechanical forces on the skull [12]. Notably, the smaller braincases found in the urban environment differ from the responses of other small mammals which show increases in braincase size [28]. Nonetheless, future work should focus on determining variation in the relative proportions of soft tissues (muscle, brains) to more precisely determine functional differences and potential adaptations among urban and rural populations. However, it would also be useful to further explore morphological variation in three-dimensions to gain further insights into masseter function. This could be indicated through how the zygomatic arches still show reductions in urban populations from a different perspective than we found. For sex, we found strong patterns of shape divergence between males and females. A more shortened, robust skull was present in females, whereby the zygomatic region was greatly reduced relative to males, which possessed a larger, more protruding, squamous temporalis and thus larger distances between the zygomatic arch and frontal bone. Notably, males displayed more elongated snouts, with reduced crania (figure 4). In relation to the patterns seen between urban and rural habitats, this suggests that females are better adapted to the potential demands of an urban environment. Indeed, selection may be stronger on females as during parental care periods female red foxes visited dens more frequently and for longer periods of time than males. This suggests that they engage with local foraging conditions more intensively relative to males, especially given the greater caloric demands placed on them during parental care [47]. This may also confer greater cognitive demands in females explaining their relatively enlarged crania. In contrast, male red foxes engage in vigilant behaviours more frequently during periods of parental care and this may involve defensive actions that favour the faster more elongate jaws we observed. If selection is driving a stronger evolutionary response to urban environments in females it could lead to an overall ‘feminization’ of urban populations through sexual conflict. If so, this would also be consistent with expected changes under domestication and deserves further attention. Figure 4. Sex-based differences in the skull shape of red foxes (Vulpes vulpes) from dorsal and ventral aspects. Trends in shape variation are magnified 3× to enhance the interpretation of shape variation.
For both dorsal and ventral aspects across the Vulpini, PC1 characterized morphological change that involved lengthening and shortening of the snout that was, respectively, concomitant with a lateral widening or narrowing of the skull. Widening was especially noticeable in the dorsal view around the zygomatic arch. On PC2, it appeared that rostral width, sagittal crest length and the length of secondary palate changed together. Within Vulpini it appeared that red foxes occupied a central region of morphospace in close proximity to basal species (figure 5). We also found low levels of phylogenetic signal in both our ventral and dorsal view PC score data. In both aspects we found no evidence to suggest K significantly differed from our null hypothesis of 0 (dorsal aspect: K.mult = 0.461, p = 0.356, power = 0.98; ventral aspect: K.mult = 0.545, p = 0.129, power = 0.93), indicating both aspects lack phylogenetic signal. We note our values of K were robust to the effects of low sample size as our analysis exhibited high power to detect differences among models. Interpreting which evolutionary processes may have led to Vulpini exhibiting low phylogenetic signal is difficult, as the relationship between K and a number of these evolutionary processes, such as the rate of morphological evolution, genetic drift, or gene flow, is often complex [40]. Nonetheless, this finding in support of no phylogenetic signal allowed us to readily compare trajectories of macro- and microevolutionary divergence (pending evidence of heritable variation) directly from unmodified landmark data. Figure 5. Scatterplots depicting the morphospace of the genus Vulpes and basal species (O. megalotis, N. procyonoides). The first two principal components from the dorsal (a) and ventral (b) aspects of fox skulls are portrayed with the associated shape changes for the extremes of each axis also being depicted as deformation grids.
We found that the pattern of divergence between urban and rural habitats did not differ from the major axis of variation found in Vulpini. Specifically, observed vector values did not exceed bootstrapped confidence intervals produced from both the dorsal and ventral aspects (90° and 97° for observed angles for dorsal and ventral aspects respectively, CIs = 177° to 48°, and 108° to 32° for dorsal and ventral aspects, respectively). These results were affirmed by our additional RRPP approach, which also indicated that the magnitude of morphological change on shared trajectories was greater across the clade than between urban/rural habitats (both p > 0.001). Therefore, coupled with a lack of phylogenetic signal, our data suggest that the directions of evolution available to red foxes diverging between urban and rural habitats are not constrained by evolutionary history, yet they follow the same pattern as their clade. Specifically, in much the same way as divergence in red foxes between urban and rural habitats, Vulpini is mainly characterized by lengthening and shortening of the snout (figure 5). Thus, the conditions presented by recent anthropogenic habitats may favour phenotypes that play a role in the speciation of Vulpini. However, this could also indicate that developmental biases common to Vulpini are playing a role in determining phenotypic variation for contemporary evolution [9,13,48]. Future approaches implementing higher resolution 3D morphometrics could help to clarify this evidence by providing more comprehensive information about shape variation. Notably, some of the craniofacial features that differ between urban and rural habitats are also similar to the effects of ‘domestication syndrome’ [5,6]. These can include traits such as docile behaviour, craniofacial morphology, ear floppiness, reductions in brain size, reduced sexual dimorphism and changes in pigmentation [9,48]. Specifically, in red foxes experimental domestication via selection on behavioural traits, more precisely ‘friendliness’ towards humans, has resulted in reduced muzzle and jaw sizes that accompany docility [15]. While not domesticated, urban foxes show reductions in muzzle size, reduced sexual dimorphism, and a narrowed braincase, and it is plausible that taking up residence in the presence of humans would favour individuals with reduced levels of fear and stress (i.e. urban tameness) as it has in other animals [49]. Mechanistically in experimental domestication this could be traced to a reduced size and function of the adrenal glands, but how could this relate to changes in the craniofacial apparatus? Recently, Wilkins et al. [9,48] proposed a link among ‘domestication’ traits that can be traced to neural crest cells (NCCs). NCCs are a vertebrate-specific class of stem cells that first appear during early embryogenesis at the dorsal edge (crest) of the neural tube. These cells migrate throughout the body toward the cranium and trunk and provide the cellular precursors of many cell and tissue types, including many of the bony elements of the skull, and the adrenal medulla [50]. In line with the idea that changes in NCCs underlie these traits, a number of neural-crest-related genes have been implicated in domestication processes [48,51]. Our findings of craniofacial divergence along an urban/rural habitat axis in red foxes suggests that some phenotypic traits related to domestication are involved, and perhaps influenced by developmental bias present within Vulpini that generally funnels variation toward a long/short jaw axis [12,13,52]. While these differences may be adaptive, they could also arise from founder effects, or other random processes. Additionally, the urban environment may actually relax selection, if it provides greater food resources and a reduced need to hunt. The inclusion of additional urban/rural gradients in future studies would be useful for discerning these possibilities. Regardless, it is notable that the trajectory of divergence taken by red foxes in response to urbanization is similar to that found over the past 15 Myr of fox evolution, suggesting that biases could be having long-term effects. Phenotypically, an interesting next step would be to assess the heritability of the morphological traits we characterize here along with a suite of additional traits related to domestication. For example, and following that these patterns mirror changes during domestication, it could be possible to experimentally compare responses to behavioural stress between fox populations in these habitats and determine whether this corresponds with variation in the size of the adrenal gland. However, genomic analyses could also enable a wide array of traits to be implicated as an evolutionary factor and identified for further study. The recently assembled red fox genome has already been used to identify regions associated with tame and aggressive behaviours [51]. Genes related to neural crest activity are well characterized [50,53] and would provide an interesting inroad into the mechanisms underlying adaptive divergence in anthropogenic environments, initiators of domestication, as well as long-term macroevolutionary change. Ventral and dorsal landmark data for Vulpes vulpes. Ventral and dorsal landmark data for genus and ancestors. Data available from the Dryad Digital Repository: https://doi.org/10.5061/dryad.bnzs7h47c [54]. K.J.P. conceived the ideas, collected data, analysed the data and led the writing of the manuscript. A.R. and H.Z. collected data and assisted with writing and analysis. A.J.C. edited drafts and performed analyses while A.K. provided access to samples, information about them, and helped by editing drafts. S.H. donated samples to the NMS collection. K.J.P., A.R., H.Z. and A.K. contributed critically to the drafts and gave final approval for publication. We declare we have no competing interests. During the time of writing K.J.P. was supported by a grant from NERC (NE/N016734/1). We thank Zena Timmons and Roberto Miguez for assistance with museum collections in Edinburgh and London respectively. We thank Neil McLean for photos of the red fox skull in figure 1. We also thank W. J. Cooper for providing photos of kit foxes. The input from three anonymous reviewers greatly improved this manuscript. FootnotesElectronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.4977716. © 2020 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original author and source are credited. References
Page 16Historical events and population interactions have shaped the differences among the over 7000 languages spoken in the world today. Genetics strongly influences functions essential for oral communication, articulation [1], hearing, [2] and phonological processing [3]. Phonological processing is the ability to identify meaningful information in a stream of speech sounds, which requires faithful spectral and temporal encoding in the auditory cortex [4]. There are several genes that modify this encoding, including DCDC2, which we previously showed has population effects on phoneme inventories [5]. DCDC2 is associated with reading and language performance and related disorders of dyslexia, specific language impairment, and speech sound disorders [6]. READ1, a regulatory element encoded in the second intron of DCDC2, is also associated with normal variation in phonological processing [7]. It has over 40 alleles that differentially enhance transcription through the DCDC2 promoter [8], which can be divided into three groups based on the presence (RU1-2) or absence (RU1-1) of an ancient 13 base duplication, or a 2.4 kb microdeletion encompassing the entire READ1 sequence. We previously showed that the frequency of RU1-1 alleles correlates with the number of consonants in the languages of 43 populations from five continents and major language families [7]. Functionally, in rodent models Dcdc2 modifies speech-sound discrimination between consonants [9,10] and is critical for temporal precision of stimulus-driven action potential firing and baseline excitability in neurons of the auditory cortex [11,12]. These studies suggest that through their effect on DCDC2 transcription, some READ1 alleles can enhance temporal precision and speech sound discrimination to favour retention or acquisition of selected consonantal contrasts. In the present study, we build on and expand our previous results to investigate the relationship between the population frequency of RU1-1 and the numbers of specific manners of consonants used in a language. Consonant manners differ in their temporal and spectral characteristics. The perception of these different manners may be more or less sensitive to the differences in temporal processing conferred by RU1-1 specific regulation of DCDC2, favouring retention or acquisition of selected consonantal contrasts. Consonants are described by three cardinal articulatory features: manner of articulation, place of articulation, and voicing [13]. Manner of articulation is the configuration and interaction of the tongue, lips, palate, and nose when making a speech sound and is subdivided into obstruents and sonorants (table 1). Obstruents and sonorants rely on high and low acoustic frequency energy respectively. Obstruents are further divided into three manners called stops, fricatives, and affricates. Stops and affricates require a complete closure of the airway; fricatives are produced by partial closure. The high frequency energy of fricatives and affricates are more robust to noise in phonological processing, whereas the transient release bursts of stop and affricates are less robust [14]. The two sonorant manners are nasals and approximants. All of these manners of articulation result in acoustic spectrograms that can be recorded from the auditory cortex as distinct neurograms [15].
We hypothesize that the acoustic differences among consonants create perceptual challenges that may make some consonants more vulnerable to loss of temporal precision than others [14] and thus more likely to be associated with RU1-1 alleles. We modelled association between RU1-1 alleles and manner of articulation in 51 populations, spanning five continents and all major language families while accounting for geographical proximity, and genetic and linguistic relatedness. For the current analysis we added nine populations from the 1000 Genomes Project (1KG) [16] to the 43 populations we used in our previous study (electronic supplementary material, table S1). We then correlated the number of consonants and vowels used in each language against the frequencies of three groups of READ1 alleles: RU1-1, RU1-2, and the READ1 2.4 kb microdeletion. Using this expanded set of populations showed that the number of consonants, but not vowels, correlated with the frequency of RU1-1 (r = 0.314, p = 0.025) in their respective population, replicating our previous results. To identify the specific linguistic features that contribute to this association, we then correlated the consonants that comprise the five manners of articulation with the frequency of RU1-1 (table 2) in all 51 populations. The number of stops had the strongest association at r = 0.398 (p = 0.0038), however RU1-1 frequency and numbers of stops clustered by regional location (figure 1). We, therefore, controlled for the effects of genetic relatedness, geographical proximity, and linguistic relatedness between populations (see electronic supplementary material for details). For genetic relatedness, 164 informative single nucleotide polymorphisms (SNPs), independent from RU1-1, were used to compute pairwise Fst values between populations. For geographical proximity, we modelled migratory distances between populations and the putative location of human origin using great circle distances along with migratory waypoints, which increase the accuracy of these distances. For linguistic relatedness, the sound inventory of the populations was used to compute the degree of inventory overlap between populations. Figure 1. Fifty-one populations, plotted by longitude and latitude. The size of the circles is defined by the frequency of RU1-1 in that population. Circles are coloured by the number of stops. (Online version in colour.)
To better understand these three effects, we evaluated how genetic relatedness, geographical proximity, and linguistic relatedness correlate with each other using the Mantel test, a statistical test of the correlation between distance matrices. Statistical significance was determined using 1000 permutations. Genetic relatedness and geographical proximity are the most strongly correlated with r = 0.563 (p = 0.001), while linguistic relatedness does not correlate with either genetic relatedness (r = −0.0771, p = 0.860) or geographical proximity (r = 0.01434, p = 0.448). The result of the Mantel tests suggests that all three effects should be taken into account in the modelling of the association between the five manners of articulation and RU1-1, given that linguistic relatedness is relatively independent from geographical proximity and genetic relatedness, and geographical proximity and genetic relatedness are only moderately correlated. To control for confounding due to effects of genetic relatedness, geographical proximity, and linguistic relatedness, we created a generalized linear regression model and included the major principal components (PCs) for each of the three sets of pairwise distances between populations. Insufficient populations were available to simultaneously examine all five manners, vowels, tones, and the control variables (the PCs) in the regression model [17]. Therefore, we performed nested model comparisons with all possible combinations of the five manners and the number of vowels and tones to avoid over-fitting, and found that the most parsimonious model contained stops and nasals: RU1-1∼stops+nasals+genetic PC-1+genetic PC-2+genetic PC-3+geographical PC-1+geographical PC-2+geographical PC-3+linguistic PC-1+linguistic PC-2+linguistic PC-3+linguistic PC-42.1 Statistical significance was estimated using 10 000 permutations. RU1-1 was significantly associated with stops and only nominally associated with nasals, but in opposite directions (table 3 and figure 2). To evaluate whether the associations were robust across different population subsets we performed a series of leave-k-out analyses by population from k = 1 (jackknife) to k = 30 (electronic supplementary material, tables S4). The directions of the beta-values remained positive for stops and negative for nasals for 95% of the subsets even when almost 50% (k = 26 for stops and k = 23 for nasals) were excluded, showing that the associations were not driven by particular subsets.Figure 2. The relationship between RU1-1 frequency and the number of consonants by stops and nasals (log-transformed, z-scores) as fitted in the best regression model (equation (2.1), table 3). Shaded regions are 95% confidence intervals. (Online version in colour.)
Our results revealed that RU1-1 is associated positively with stops and negatively with nasals in 51 populations, adjusting for geographical proximity, and genetic and linguistic relatedness. To contextualize our results, we examined consonant processing in published animal models. Recordings from the auditory cortex of wild-type rats in response to different consonant manners showed that neurograms from stops have the sharpest onset peaks, are the most distinctive manner of articulation, and can be distinguished better than other manners in the first 40 milliseconds (ms) [9]. By contrast, while it takes longer to discriminate between pairs of nasals (greater than 40 ms), they are more robust to loss of millisecond temporal precision compared to stops, which are more sensitive. Disruption of Dcdc2 in rodents also supports a prominent role in differential perception of stops and nasals. RNAi knockdown (KD) of Dcdc2 in rats leads to an increase in excitability in neurons located in the auditory cortex, as evidenced by both spontaneous and stimulus-driven action potentials and shorter onset latency in firing [18]. The correlate in the Dcdc2 knockout (KO) mouse also shows an increase in neuronal excitability demonstrated by an elevated resting membrane potential in whole-cell patch clamp studies of brain slices [11]. KO mice also have decreased temporal precision of stimulus-driven action potential firing. This decrease is mediated by increased expression of Grin2B, a subunit of the N-methyl-D-aspartate receptor (NMDAR), and restored by an NMDAR antagonist. Behaviorally, Dcdc2 KD rats have defective consonant detection in speech streams presented at all speeds, and altered neurograms at very high speeds recorded from the auditory cortex. Dcdc2 KO mice have impaired rapid auditory processing [12]. These studies show that in rodent models, Dcdc2 is important for potentiating baseline excitability and temporal precision in neurons of the rodent auditory cortex, which are critical for the sensitive and accurate perception of specific sound targets, and consonants in particular. Human DCDC2 promoter reporter-gene constructs show that the most frequent RU1-1 allele confers higher expression than the most frequent RU1-2 allele [8]. This suggests that RU1-1 alleles increase DCDC2 expression, and as suggested by the rodent experiments, could lead to increased temporal precision in the auditory cortex and enhanced consonant discrimination (figure 3). Figure 3. Experimental lines of evidence for the role of DCDC2/RU1-1 in the number of consonants in a language. Dcdc2 KD rat has reduced consonant discrimination in a stream of speech sounds. Dcdc2 KO mouse has reduced temporal precision. Created with biorender.com. (Online version in colour.)
Together, these data suggest that RU1-1 alleles acting through increased expression of DCDC2, increase auditory processing precision that enhances stop-consonant discrimination. In this model, populations with higher RU1-1 allele prevalence would have enhanced ability to discriminate between similar stop-consonants, and over time, change the phoneme inventory of the language by addition or retention of words with similar stops, expand the stop-consonant repertoire, and increase the total number of stop-consonants relative to other sounds. When RU1-1 prevalence is low in a population, fewer individuals would have the ability to finely discriminate between stops. Thus, stop consonants would potentially be fewer, and nasals, which do not require enhanced temporal precision and are robust to noise, would be more likely to be recruited or retained. This would account for the inverse relationship of stops and nasals that we observed (figure 2). Linguistic theory holds that listeners play an important role in shaping the sound structure of language [19,20] through perceptual biases that introduce errors [21], and result in vocal compensation to ensure effective communication [22,23]. For example, a high prevalence of chronic ear infections (up to 90%) in some Australian Aborigine populations is thought to be the cause of a phonemic inventory lacking sounds that depend on acoustic frequencies impacted by the infections [24,25]. Differences in the auditory processing of stops and nasals are more subtle than loss of acoustic frequencies, but nonetheless have physiologic correlates in evoked response potentials measured in the auditory cortex [15]. While the results of the current analyses show a significant association of stops and nasals with RU1-1 at the population level, it is important to note that subjects were not phenotyped individually for differences in speech perception. Additionally, language assignment has potential for errors and phonemic inventory counts can differ between studies and sub-populations. While our genetic samples were chosen to be representative of the population as a whole, hidden population stratification not accounted for by the PC corrections could distort the findings. Although the overall effect size of RU1-1 on language change is likely to be small, subtle cognitive biases can be amplified through cultural transmission to shape the structure of languages over time; simulated models of evolution show that a small difference (as little as 5%) in population prevalence of a bias in favour of a linguistic change can disproportionately increase the number of speakers by up to 45% [26]. In addition, the link between DCDC2 expression and stop-consonant discrimination relies on published experiments in rodents. Further studies in humans are necessary to demonstrate a more direct genotype–phenotype connection. It should also be noted that although we interpret the association between RU1-1 and stop-consonant description as a possible driver for phoneme inventory change among populations, a strict interpretation of the analysis means that the inverse relationship is also possible and that certain phoneme inventories may have favoured genetic selection. However, we view this as unlikely, given arguments from computational simulations of language change and experimental evidence from human artificial language learning and song birds [27]. Language is continuously evolving to meet the needs both of the speaker and of the listener [28]. The needs of the speaker include balancing ease of articulation and communicative success. The needs of the listener include ease of decoding by having a signal that is robust to noise. Stop consonants perform well when background noise is low and listeners with RU1-1 alleles have greater capacity to discriminate between them. When background noise is prominent, listeners without RU1-1 alleles may have reduced capacity to discriminate between stops—nasals are preferred because of their robustness to noise. The nature of these consonant manners and the direction of their relationships with RU1-1 supports an account of how languages evolve to maintain phonemes that are robust to auditory processing constraints [20]. These findings enhance classical linguistic theories on the evolution of language shaped by historical migrations [29,30], conquests [31], admixtures [32], geographical isolation [33–37], diet [38], and communication efficiencies [28,39] by adding a genetic dimension, which until recently, has not been considered to be a significant catalyst for language change. All data and our analyses are available in the electronic supplementary materials and are available via the Dryad Digital repository https://dx.doi.org/10.5061/dryad.hdr7sqvf4 [40]. Conceptualization, K.T., M.M.C.D., and J.R.G.; investigation, K.T., M.M.C.D., and J.C.F.; writing—original draft, K.T and M.M.C.D.; writing—editing, K.T., M.M.C.D., and J.R.G.; funding acquisition and supervision: J.R.G. Yale University has applied for a patent covering READ1. J.R.G. is an inventor named on the patent application. Support for J.R.G., M.M.C.D., and J.C.F. was provided by The Manton Foundation. Support for J.R.G. was provided by NIH/Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD) P50 HD027802. We wish to thank Dongnhu Truong, Jeffrey Malins, and Andrew Adams for helpful discussions. We thank Bill Speed for providing a merged dataset of Kidd laboratory and 1KG SNPs and Julian DeMille for his assistance with data extraction. Footnotes†Both authors contributed equally. Electronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.4977686. © 2020 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original author and source are credited. References
Page 17Predator–prey interactions are ubiquitous across ecosystems. Predation has been widely studied at an ecological level [1–3], and recent research also shows that this interaction can be strongly altered by rapid evolution of anti-predatory defence in the prey [4] as well as by counter-adaptations in the predator [5–7], even though selection may be asymmetric, resulting in slower evolutionary change for the predator [8]. Moreover, owing to population growth–defence trade-offs, rapid evolution of the prey and adaptation to predation can result in frequency-dependent selection of defended and undefended prey types as a function of predator population size [9–11], an example of eco-evolutionary feedback dynamics. Common to this spectrum of evolutionary, coevolutionary and eco-evolutionary dynamics is that these dynamics are all driven by natural selection acting on fitness-relevant traits. Predation can be described by three main phases, namely prey search, capture and ingestion [12]. These three phases are shaped by key traits in predator–prey systems, including those influencing offence and defence level, and all these traits can be subject to evolutionary change [13]. The offence level is determined by sensory faculties and speed enabling location and capture of prey, and defence level by the capacity for predator avoidance and escape prior to ingestion as well as physico-chemical obstruction of ingestion and digestion [12]. Adaptations in defence and offence, in turn, combined with associated trade-offs, modulate the reproduction (i.e. life-history traits) of both parties [14]. Examples abound of the study of the different phases of predation, and adaptation in both predator and prey life-history traits. For example, the timing and population dynamics of many insectivorous bird species are tightly coupled to the dynamics of their prey insect species [15]. Olive baboon sleeping site choice and behaviour (sharing sleeping sites between multiple baboon groups) in Kenya were recently linked to decreased contact and capture rate by leopards [16]. Coevolution has been hypothesized to occur between northern Pacific rattlesnakes and California ground squirrels whereby venom resistance in squirrels is matched by increased venom effectiveness in rattlesnakes based on field data supportive of local adaptation of the traits [17]. The empirical examples of evolving predator–prey interactions described above cannot be used to experimentally investigate (co)evolution in predator–prey systems owing to the long generation times of the species. By contrast, microbial systems offer a unique opportunity to study predator–prey dynamics, as they include efficient (high prey capture rate) predators and allow for high replication as well as experimental approaches capturing both ecological and evolutionary dynamics. Microbial predator–prey systems show many key characteristics found also in other predator–prey systems, such as offence by speed [18] and defence by avoidance of detection [19], escape [20] or physico-chemical obstruction of ingestion or digestion (for an overview, see [12]). Defence level has also been demonstrated to evolve in controlled set-ups [21,22]. However, to our knowledge, there exist little to no empirical studies examining offence mechanisms subject to rapid evolution in microbial predator–prey systems. Here, we employed an experimental evolution approach to test the influence of approximately 600 generations of predator–prey interaction on predator traits, using a microbial (ciliate–bacteria) model system. Since predator–prey dynamics are characterized by the intrinsically linked dynamics of both interaction partners, we inspected the influence of both prey and predator evolution on predator traits. To find general patterns in predator traits independently of any specific prey species, as most predators have multiple prey species [23], we used seven different prey species that were all separately evolved with the predator. We expected rapid evolution of anti-predatory defence in the prey to cause impairment of predator growth [7,14]. We expected predator evolution to be weaker, in line with the life–dinner principle [8,24] positing that the prey experiences stronger selection pressure since its survival (life) directly depends on defence, while the predator can afford a certain measure of unsuccessful prey encounters (dinner postponement). Asymmetric selection can result in dynamics other than classic arms race dynamics such as frequency-dependent cycling of traits [5], which have also been observed in microbial predator–prey systems [22]. Nevertheless, instead of escalation where predators alone impose selection pressure, we expected to also observe predator evolution, since coevolution has been demonstrated to occur in bacteria–ciliate systems, in line with the Red Queen hypothesis [7,14,25]. We studied the evolutionary dynamics of one focal predator species (the ciliate Tetrahymena thermophila) and seven of its bacterial prey species in all seven combinations of predator–prey species communities, as well as dynamics in prey species populations only. We ran predator–prey evolutionary experiments over about 600 predator generations, and assessed evolutionary effects on life history, morphology and behaviour using common garden experiments. The seven prey species used in this study are listed in table 1. In addition to four taxa previously used as models in predator–prey studies, three strains were chosen based on representing genera associated with ciliate predators in natural habitats or potentially exhibiting different anti-predatory defence mechanisms (table 1). Since each strain represent a single genus, strains are referred to by their genus names in the text.
We used a single strain of the asexually reproducing ciliate T. thermophila 1630/1U (CCAP) [31] as a generalist predator capable of consuming all the prey species. Tetrahymena thermophila is a ciliate species characterized by a facultative sexual reproductive cycle and nuclear dualism, where the cells contain a small diploid non-expressed germline nucleus (micronucleus) and a larger highly polyploid somatic nucleus (macronucleus), derived from the micronucleus after sexual reproduction [32]. Only the macronuclear DNA is expressed and hence determines the phenotypic characteristics of Tetrahymena cells [32]. The micronucleus is only relevant for sexual reproduction. The species can be maintained either under settings of recurrent sexual reproduction, or as asexual lineages only. The Tetrahymena strain used in our experiment had been maintained in serial propagation for many years before the experiments. Sexual reproduction only occurs when induced by starvation [33], and because this was not the case during its long-term maintenance, the strain only underwent asexual reproduction. During asexual reproduction, micronuclei and macronuclei divide independently from each other [32]. It has been noted that, when cultured for a long time asexually, the micronuclei can degrade [34] and have subsequent negative effects on the genotype's fitness during a possible sexual reproduction, or even lead to genotypes losing their ability to sexually reproduce. However, given that micronuclei are never expressed and only play a role in sexual reproduction [32], and also given that we do not induce or study the genotype's ability to reproduce sexually, this possible degradation of the micronucleus does not have consequences on fitness as measured in our setting. We also note that it is a common practice to use Tetrahymena cell lines with non-functional micronuclei, as described in the standard handbook for Tetrahymena cell biology work [34]. In all of these cases, the serial propagation is not problematic as long as one is not inducing sexual reproduction. Hence, any evolution observed at the predator level in this experiment stems from either mutations or selection on existing variation in the macronuclear DNA. Furthermore, as the macronucleus is highly polyploid (n = 45), and chromosomes divide randomly during asexual reproduction [32], cells are relatively buffered to the effects of single maladaptive mutations, and can undergo relatively rapid purging of maladaptive mutations or selection for increased copies of adaptive mutations. This, together with the absence of sexual reproduction, which can be affected by serial propagation [34], makes it highly unlikely that the serial propagation set-up in the experiment would itself strongly influence the evolutionary dynamics of the predator. Prior to the experiments, all bacterial stocks were kept at –80°C and ciliate stocks were cultured axenically in proteose peptone yeast extract (PPY) medium containing 20 g of proteose peptone and 2.5 g of yeast extract in 1 l of deionized water. During the evolutionary experiment, cultures were kept at 28°C (±0.1°C) with shaking at 50 r.p.m. The evolutionary experiment was started using a small aliquot (20 µl) of a 48 h bacterial culture started from a single colony and 10 000 ciliate cells (approx. 1700 cells ml–1) from an axenic culture. Each bacterial strain was cultured alone and together with the ciliate predator (three replicates each, with the exception of six replicates for Comamonas) in batch cultures of 20 ml glass vials containing 6 ml of 5% King's B (KB) medium, with 1% weekly transfer to fresh medium. Every four transfers (28 days), bacterial and predator densities were estimated using optical density (1 ml sample at 600 nm wavelength) as a proxy for bacterial biomass and direct ciliate counts (5 × 0.5 µl droplets using light microscopy) as used in this context and described previously [30,35,36], and samples were freeze-stored with glycerol at –20°C for later analysis. Since predators do not survive freeze-storage in these conditions, at time points 52 and 89 weeks, predator cultures were made axenic by transferring 400 µl into 100 ml of PPY medium containing an antibiotic cocktail (42, 50, 50 and 33 µg ml–1 of kanamycin, rifampicin, streptomycin and tetracycline, respectively) and stored in liquid nitrogen. Axenicity was controlled for by plating on agar plates containing 50% PPY medium, on which all the experimental bacterial strains grow. The liquid nitrogen storage protocol was modified from a previously used protocol [34] and included starving a dense ciliate culture in 10 mM Tris-HCl solution (pH 7) for 2–3 days, centrifugation (1700g, 8 min, 4°C), resuspension of the pellet in 1 ml of leftover supernatant and the addition of 4 ml of sterile 10% dimethyl sulfoxide (DMSO). The resultant solution was transferred to cryotubes in 0.3 ml lots, and frozen in a –20°C freezer at a rate of –1°C min−1 using a Mr Frosty™ Freezing Container (Thermo Scientific) for cell preservation before transferring to liquid nitrogen. We isolated the populations for the current experiment at time point 89 weeks (approx. 20 months). With the minimal assumption that populations multiply by 100-fold (dilution rate) until reaching the stationary phase, each weekly transfer interval represents 6.64 generations for both prey and predator [37], constituting a total minimum of approximately 600 generations. Community dynamics are shown in electronic supplementary material, figures S1 and S2 and demonstrate clear differences in population size between different prey species. Bacteria were restored from freeze-storage by transferring 20 µl into 5 ml of 5% KB medium and culturing for 72 h. Predators were restored from liquid nitrogen by thawing cryotubes in a 42°C water bath for 15 s, followed by the addition of 1 ml of 42°C PPY medium. The cryotube contents were then transferred to a Petri dish containing PPY medium at room temperature. Upon reaching a high density (approx. 48 h), predators were transferred to 100 ml of PPY medium and cultured to a high density (approx. 7 days). To ensure that the antibiotic treatment or the liquid nitrogen storage and revival procedures do not contribute to potential differences between the ancestral predator and evolved predator lines, the axenic ancestral predator was subjected to identical procedures and was revived at the same time as the evolved lines. These culturing steps representing over 10 generations should remove the influence of non-genetic changes in predator traits caused by phenotypic plasticity [38]. To test bacterial and ciliate performance and traits, we used a combination of automated video analysis, optical density measurements and flow cytometry. To separate evolutionary responses at the predator and prey level, we tested performance of both evolved and ancestral bacteria with evolved and ancestral ciliates for all evolved lines reciprocally. To do so, we prepared 12 50 ml Falcon® tubes by adding 20 ml of 5% KB medium. Three of these were inoculated with ancestral bacteria and ancestral ciliates, three with ancestral bacteria and evolved ciliates, three with evolved bacteria and ancestral ciliates and the remaining three with evolved bacteria and evolved ciliates. We placed the Falcon® tubes in a 28°C incubator, rotating on a shaker at 120 r.p.m. After inoculation, the samples were left to grow for a period of 12 days, to allow populations to grow to equilibrium density. Over the course of these 12 days, we took a total of 10 samples from each culture for analysing population density dynamics of bacteria and ciliates, and morphological and behavioural metrics for the ciliates. We sampled cultures by gently shaking the culture, to ensure it was well mixed and subsequently pipetting out 200 µl from the mixed culture. Bacterial density was determined both through measurement of optical density and through flow cytometry. Flow cytometric analyses were based on established protocols [39,40] that facilitate distinction between living bacterial cells and background signals (e.g. dead cells or abiotic matter). For flow cytometry, we sampled 50 µl of all cultures, diluted the samples 1 : 1000 using filtered Evian water and transferred 180 µl of the diluted samples to a 96-well-plate. We then added 20 µl of SybrGreen to stain the cells and measured bacterial cell counts using a BD Accuri™ C6 flow cytometer. As the inner diameter of the needle from the flow cytometer was 20 µm, and hence smaller than typical ciliate cell sizes, it is highly unlikely that ciliate cells were accidentally measured during flow cytometry. Also, given that bacterial densities were typically between one and five orders of magnitude larger than ciliate densities, even an occasional measurement of ciliate cells would have a negligible effect on bacterial density estimates. The full protocol can be found in the electronic supplementary material. For optical density measurement, we sampled 50 µl of all cultures, diluted 1 : 10 using filtered Evian water, and measured absorbance at 600 nm using a SpectroMax 190 plate reader. For measuring ciliate density, we performed video analysis [41] using the BEMOVI R-package [42]. We followed a previously established method [43] where we took a 20 s video (25 frames s−1, 500 frames) of a standardized volume using a Leica M165FC stereomicroscope with circular lighting and mounted Hamamatsu Orca Flash 4.0 camera. We then analysed the videos using BEMOVI [42,44], which returns information on the cell density, morphological traits (longest and shortest cell axis length) and movement metrics (gross speed and net speed of cells, as well as turning angle distribution). The video analysis script, including used parameter values, can be found in the electronic supplementary material. All statistical analyses were done using the R statistical software (v. 3.5.1) [45]. To obtain the reported F- and p-values for predator traits, we performed ANOVA for the best linear models constructed for the different traits as described below. To visualize whether the full set of trait data displayed structure depending on the evolutionary history of the predator and prey species, t-distributed stochastic neighbour embedding (t-SNE) was performed for each prey species separately using the Rtsne package [46] with a perplexity parameter of 3 owing to small sample size. For analysing the population growth dynamics of the ciliates, we implemented the Beverton–Holt population growth model [47] (electronic supplementary material, figure S3) using a Bayesian framework in RStan [48], following methods used by the authors in [49,50] . This function has the form of dNdt=(r0+d1+αN−d)N, with r0 being the intrinsic rate of increase, α the intraspecific competitive ability and d being the death rate in the population. Model code for fitting this function can be found in a Github repository (doi:10.5281/zenodo.2658131). For fitting this model, we needed to provide prior information for r0, d and equilibrium density K. The intraspecific competitive ability α was later derived from the other parameter values asα=r0Kd. The priors (lognormal distribution) of the model were chosen in such a way that the mean estimates lay close to the overall observed means, but were broad enough so the model was not constrained too strongly:
Models were run with a warm-up of 2000 iterations and a chain length of 8000 iterations. We analysed the estimates of the life-history traits obtained from the Beverton–Holt model fit (r0, α and K) using linear models and model selection. We first constructed a full model with life-history traits being a function of bacterial evolutionary history (evolved/ancestor), ciliate evolutionary history (evolved/ancestor) and bacterial species (seven species factors) in a full interaction model. Next, we used automated bidirectional model selection using the step function (stats package v. 3.5.1) to find the best model. To avoid bias due to starting point, we fitted the model starting from both the intercept model and the full model, and if model selection resulted in different models, we used sample-size adjusted Akaike Information Criterion (AICc) comparison (MuMIn R-package, v. 1.42.1 [51]) to select the model with the smallest AICc value. Morphological and behavioural data were available for every time point during the growth curve, and since we know these traits can be plastically strongly affected by density [52,53], we had to take density into account in the model. We hence separated the analysis into two steps: first, we identified key points in the growth curves (early phase, mid-log phase and equilibrium density phase) and analysed the traits for these particular points. Secondly, we fitted models over all data, but taking bacterial (using flow cytometry data) and ciliate densities into account as covariates in the statistical analysis. We defined the early phase as the second time point in the time series, equilibrium density phase as the first time point where density was larger than 99% of K, or alternatively the highest density, and the mid-log phase as the point between the early and equilibrium density phase where density was closest to 50% of K. We then created statistical models for the traits (major cell axis size, gross speed of cells and turning angle distribution) as a function of bacterial evolutionary history (evolved/ancestor), ciliate evolutionary history (evolved/ancestor) and bacterial species (seven species factors), including a full interaction for the data at the particular time point. Next, we used automated bidirectional model selection to find the best-fitting model. This was done separately for all three phases (early, mid-log and equilibrium density phases). We again performed model selection starting from both the intercept model and full model, and compared the two models using AICc comparison to identify the best model. We then created models using all the data, where we fitted major cell axis size, gross speed and turning angle distribution as a function of bacterial evolutionary history (evolved/ancestor), ciliate evolutionary history (evolved/ancestor) and bacterial species (seven species factors), ciliate population density (ln-transformed, continuous) and bacterial population density (ln-transformed, continuous), including a full interaction. For turning angle, we also did a log10 transformation of the turning angle distributions, as fitting the model on untransformed data leads to a strong deviation on the qqplot. Next, we used automated bidirectional model selection using the step function starting from intercept model and full model, and compared the two models using AICc comparison to select the best model. The t-SNE maps (figure 1) showed that the evolutionary history of the predator and prey species frequently resulted in predator divergence in trait space. Importantly, this divergence evolved from a single ancestral predator population, which was subjected to co-culture with different prey species. The full results for all statistical analyses presented below to assess this divergence in detail are available in the electronic supplementary material. Figure 1. t-SNE map of contribution of predator and prey evolutionary history to predator divergence in trait space. The traits included in the analysis encompass life history (intrinsic growth rate, equilibrium density and competitive ability), morphology (cell size and biovolume) and behaviour (speed and cell turning angle distribution). (Online version in colour.)
Prey evolution drove changes in the life-history traits of the predator, including intrinsic rate of increase (r0), equilibrium density (K) and competitive ability (α), although the presence and strength of the effect depended on the bacterial species (ANOVA, r0: prey evolution F1,78 = 15.32, p < 0.001; prey evolution × prey species F6,78 = 9.03, p < 0.001; K: prey evolution × prey species F6,80 = 13.7, p < 0.001; α: prey evolution F1,78 = 4.79, p = 0.031; prey evolution × prey species F6,78 = 5.40, p < 0.001; electronic supplementary material, tables S1–S3 and S7–S9; figure 2). The intrinsic rate of increase of ciliates (r0) was generally lower in the presence of evolved bacterial prey compared with ancestral prey, with the notable exception of Serratia, where intrinsic rate of increase was higher in the presence of evolved prey (table 2 and figure 2). For three species (Brevundimonas, Janthinobacterium and Pseudomonas), evolved predators had a higher intrinsic rate of increase (r0) on evolved prey compared with ancestral prey (figure 2). Changes in population equilibrium density (K) were highly dependent on species, with four species (Brevundimonas, Comamonas, Janthinobacterium and Serratia) showing higher population equilibrium density in the presence of evolved prey compared with ancestral prey, and the remaining three (Escherichia, Pseudomonas and Sphingomonas) showing decreased population equilibrium density in the presence of evolved prey compared with ancestral prey. Competitive ability (α) typically decreased in the presence of evolved prey compared with ancestral prey, with the exception of Pseudomonas, where competitive ability was higher in the presence of evolved bacteria compared with ancestral bacteria. Notably, for Escherichia, Janthinobacterium and Serratia, the competitive ability (α) of evolved predators was higher in the presence of evolved prey compared with ancestral prey (figure 2). Figure 2. Reaction norms showing the effect of evolving predator–prey interaction on life-history traits of predator (data points with linear model estimate ±95% confidence intervals; N = 3, except 6 for Comamonas). The life-history traits for predators are parameters of Beverton–Holt continuous-time population models fitted to data, and include intrinsic growth rate (r0), equilibrium density (K) and competitive ability (α). The reaction norms for predators (one strain of the ciliate Tetrahymena thermophila) feeding on ancestral or evolved prey (seven bacterial strains indicated by genus name) are depicted separately for ancestral and evolved predators (colour coding). Predators evolved with a particular prey taxon have always been coupled with ancestral or evolved populations of the same taxon, while the ancestral predator is the same for all prey taxa. (Online version in colour.)
In contrast with life-history traits, which were affected by prey evolution alone, morphological and behavioural traits of the predator were affected by predator evolution (figure 3). However, the effect size of predator evolution was also strongly dependent on predator density (for the movement metrics gross speed and turning angles) or both predator and prey density (for the biovolume metric cell size). Evolved predators were slightly but significantly larger than ancestral predators (ANOVA: predator evolution F1,767 = 7.87, p = 0.005). Although there was a significant effect indicating that this was modulated by the evolutionary history of the prey (ANOVA: prey evolution F1,767 = 4.85, p = 0.033), the associated effect size was much smaller than predator evolution. On average, evolved predators were 39.12 µm larger than ancestral predators, and predators were on average 1.629 µm smaller in the presence of evolved prey compared with ancestral prey. The effect of predator evolution also depended strongly on prey densities (ANOVA: log prey density × predator evolution F1,767 = 6.87, p = 0.009; figure 3). The strongest differences in cell size between ancestral and evolved predators were observed at low prey densities (cell sizes 1.2–1.3 times larger for evolved compared with ancestral ciliates), whereas the effects were negligible at high prey densities (approximately equal size for evolved and ancestral ciliates; electronic supplementary material, tables S4 and S10 and figures S4–S6; figure 3). Figure 3. Ratios of the predicted trait values of the linear models (cell size, gross cell speed and turning angle) for the evolved predator divided by the ancestral predator at different prey densities (5, 50 and 95% quantiles) and predator densities (5, 50 and 95% quantiles). Ratios represent how ciliate traits differ between evolved and ancestral ciliates, with a value of 1 meaning evolved and ancestral ciliates are identical, values larger than 1 meaning higher trait values for evolved strains, and values smaller than 1 higher trait values for ancestral ciliates. (Online version in colour.)
The gross movement speed of predators depended on the interplay between predator density and predator or prey evolutionary history. Evolved predators had, on average, up to 1.25 times higher speed compared with ancestral predators. However, this effect occurred for evolved predators at high predator densities, whereas at low predator densities, movement speed was approximately similar for ancestral and evolved ciliates (ANOVA: predator density F1,763 = 116.20, p < 0.001; predator evolution F1,763 = 1.90, p = 0.239; predator evolution × predator density F1,763 = 4.36, p = 0.037; figure 3). This effect was partially counteracted by prey evolution by driving speed to a lower rate at increasing predator densities (ANOVA: prey evolution F1,763 = 2.17, p = 0.141; prey evolution × predator density F1,763 = 5.46, p = 0.020). The movement speed of ciliate cells was also dependent on the identity of the prey species, with ciliates moving slower when subjected to three particular prey species (Janthinobacterium, Pseudomonas and Serratia; ANOVA: prey evolution F6,763 = 9.11, p < 0.001; electronic supplementary material, tables S5 and S11 and figure S7). Finally, predator evolution altered cell turning angle distribution across prey species such that evolved predator lines moved in straighter trajectories (ANOVA: predator evolution F1,56 = 10.15, p = 0.001). This effect was again highly dependent on predator population size, with evolved predators turning at approximately 0.92 times the turning rate of ancestral predators at low predator density, but turning equally as much at high predator density (ANOVA: predator density F1,763 = 33.90, p < 0.001; predator evolution × predator density F1,763 = 5.44, p = 0.02; figure 3). The effect of predator population size was also dependent on prey species, such that for three prey species (Janthinobacterium, Pseudomonas and Serratia), evolved predators moved even straighter (less turning) at higher predator densities (ANOVA: predator density × prey species F1,763 = 6.76, p < 0.001; electronic supplementary material, tables S6 and S12 and figures S8–S10). We quantified the contribution of predator and prey evolution to predator trait change across seven different prey species in a 20-month (approx. 600 predator generations) co-culture experiment. Prey evolution frequently led to changes in predator life-history traits, decreasing intrinsic growth rate, equilibrium density or competitive ability, while not affecting morphological or behavioural traits in the predator. Interestingly, the strength of the effect and the life-history trait affected depended on the prey species. These results may be influenced by different growth dynamics, defence levels or defence mechanisms of the different prey species (table 1; electronic supplementary material, figures S1 and S2) [12]. For two of the predator life-history traits, intrinsic rate of increase (r0) and competitive ability (α), the trait was impaired, with evolved compared with ancestral prey in all except for two cases (Serratia for r0 and Pseudomonas for α). This could be caused by any mechanism of prey defence evolution decreasing effective prey population size or increasing prey handling time, including cell aggregation of bacterial prey, frequently shown under ciliate predation [54,55]. While a similar result was also observed for population equilibrium density (K) with three prey species (Escherichia, Pseudomonas and Sphingomonas), intriguingly, the remaining four prey species (Brevundimonas, Comamonas, Janthinobacterium and Serratia) showed higher K in the presence of evolved compared with ancestral prey. This counterintuitive result may be caused by resource use evolution, which can occur rapidly in bacterial evolutionary experiments [37] but differ in magnitude between bacterial (i.e. prey) species. In this situation, a sufficient increase in prey population size could sustain a higher predator population size despite anti-predatory defence evolution. Consistent with the Red Queen hypothesis, evolved predators displayed both behavioural and morphological changes linked to prey foraging efficiency. Increased swimming speed and body size were observed for evolved predators with certain prey species, and predators evolved to swim in straighter trajectories across the different prey species. Increased swimming speed and decreased cell turning (i.e. moving in straighter trajectories) have both been linked to prey search efficiency [18,56,57], and in line with this, ciliates have been shown to display decreased cell turning and increased speed at low food concentrations [58]. The role of increased body size is less clear but may also be related to increased prey search efficiency since swimming speed can be a function of body size [18,56]. All these evolutionary trait changes in the predator are consistent with being adaptations to decreased food availability owing to anti-predatory defence evolution in the prey species. Interestingly, against our expectation based on the Red Queen hypothesis, we did not find detectable levels of adaptation in predator life-history traits when prey-evolved predators fed on their respective ancestral prey species. This could be indicative of asymmetry of selection [5,22] such that predators experience weaker selection pressure compared with prey owing to the life–dinner principle [8], whereby prey species rely on adaptation (needed to stay alive) more strongly than predators (needed to increase energy uptake). Asymmetric evolutionary change for ciliate predators could also result from smaller population size (in the order of 104 ml–1 for ciliates compared with 108 ml–1 for bacteria), larger genome size (greater than 100 Mb for T. thermophila compared with less than 10 Mb for bacteria) or more complex genomic architecture limiting adaptive mutation supply compared with the bacterial prey [59]. There are two ways asymmetric selection could account for our unexpected result regarding the lack of evolution in ciliate life-history traits. First, the offence-related traits (morphology and behaviour) where predator evolution was observed may simply not have improved sufficiently to be detectable as increased predator growth on ancestral prey using our methods. Although the culture conditions were mostly identical between the serial passage experiment and ciliate physiology measurements (same culture medium, temperature, covering 7-day time span representing serial passage culture cycle), it is also possible that minor differences in experimental conditions (different culture vials, volumes and shaking parameters) or the revival of ciliates from liquid nitrogen storage could have introduced noise in the data, masking ciliate evolution in life-history traits. Second, rapid evolution in the prey species may have changed basic features of the prey population early on in the experiment, such as causing cell aggregation, which is widely documented to evolve rapidly in similar set-ups [21,22,54,55]. An improved ability of the predator to feed on defended prey with altered characteristics may not allow for an improved ability to also feed on ancestral prey. For instance, higher speed and directionality of movement may be useful when feeding on unevenly distributed prey aggregates while not causing a benefit when feeding on prey as homogeneously distributed single cells (food being always closely available). Alternatively, as a more complex explanation, a steepening growth–offence trade-off during coevolution [14] could cause stunted growth in coevolved high-offence-level predators, which may, therefore, only display a net fitness improvement against prey in a recent evolutionary state. Since our sample material represents a snapshot from the endpoint of a long-term (co)evolutionary experiment, further experiments would be needed to assess the dynamics of predator trait change over time to test these hypotheses. Our findings have implications for interpreting data from (co)evolving predator–prey systems. First, the pronounced impairment of predator growth traits upon prey evolution together with the lack of clear improvements in the ability of evolved predators to feed on ancestral prey types support the asymmetric selection hypothesis. Second, the occurrence of predator evolution in other key traits for predator–prey interaction despite this suggests that tracking ecological changes alone may result in an underestimation of predator evolution [60,61]. A deeper understanding of predator–prey evolutionary dynamics is, therefore, likely to critically depend on the identification and examination of key traits for the interaction, preferably over time and including both interaction partners. All code and pre-processed data needed to reproduce the ecological and evolutionary analyses are available from the Dryad Digital Repository: https://doi.org/10.5061/dryad.08kprr4zr [62]. J.C. and T.H. designed the coevolutionary experiment. J.C. performed and managed the experiment and wrote the draft manuscript. F.M., E.A.F. and F.A. designed and performed physiological measurements. F.M. and J.C. analysed data. All authors interpreted results and participated in improving the manuscript. We declare we have no competing interests. This work was funded by the Academy of Finland (T.H.; project no. 106993), the University Research Priority Program (URPP) ‘Evolution in Action’ of the University of Zurich and the Swiss National Science Foundation (grant no. PP00P3_179089, to F.A.) and the Jenny and Antti Wihuri Foundation (grant no. 00190040, to J.C.). We thank Veera Partanen for technical help with maintaining the coevolutionary experiment and reviving samples for the physiological measurements. We thank Samuel Hürlemann for help during the laboratory work. We thank two reviewers for constructive comments on a previous version of the manuscript. This is publication ISEM-2020-16 of the Institut des Sciences de l'Evolution–Montpellier. Footnotes†These authors contributed equally to this work. Electronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.4970678. References
Page 18Though hybrids between mammalian species have been catalogued for decades [1], the extent and frequency of gene flow between evolutionarily divergent taxa has only been recognized since the availability of high-coverage nuclear genomes. Recent studies have revealed rampant gene flow between multiple species of bears [2], canids [3], felids [4–6], cetaceans [7,8], birds [9–11], suids [12,13] and bovids [14]. Genome analyses of invertebrate lineages including butterflies [15,16] and mosquitoes [17] have also revealed similarly extensive patterns of ancient and contemporary introgression. This demonstrated frequency of introgression is perhaps surprising given the significant barriers that maintain reproductive isolation in species pairs that diverged millions of years ago. In mammals, genomic barriers manifest in accordance with Haldane's Rule [18] as the unisexual sterility of the heterogametic sex (XY males) in F1 hybrid offspring. In cases where matings between F1s fail to produce F2s, fertile offspring can often be produced through backcrosses between the fertile F1 females and males from one of the parent species. Occasionally, however, F1s produced from interbreeding between closely related mammalian species pairs can produce viable and fertile F2 offspring. If the calculated pairwise genetic distance between two species correlated with the ability of their hybrid offspring to produce F2s, these values could serve as a proxy to predict this occurrence. Though at least one study [19] reported that genetic divergence values do generally correlate with species boundaries, others [20,21] have questioned whether this correlation exists, and have instead stated that measures of genetic distance between species are not reliable predictors of hybrid sterility. A recent empirical study of damselflies, however, demonstrated a strong correlation between the genetic distances between species pairs and their relative reproductive isolation [22]. Establishing whether genetic distance and reproductive isolation are correlated is also critical for our understanding of the genetic architecture of reproductive isolation. Doing so firstly requires knowing whether any two species are capable of producing viable or fertile offspring, but there is a general paucity of captive breeding experiments or field data that have unequivocally established this. An alternative approach is to develop a metric that can accurately predict the relative fertility of the F1 hybrids of any two species that makes use of interspecific crosses whose offspring have been reproductively characterized. Here, we developed a robust, quantitative framework based on the correlation between mitochondrial genetic distance between mammalian species known to produce F1 pairs to obtain a quantitative measure of whether F1 hybrids of both sexes are likely to be capable of breeding, or if they instead manifest Haldane's Rule. We tested the accuracy of the proxy in a well characterized felid hybrid system, and then applied it to a hominin case study to assess the relative potential sterility of hybrids between humans and their closest extinct relatives. We first explicitly defined two dichotomous categories along the spectrum of hybrid incompatibility. Category 1 is defined by mammalian species pairs capable of producing fertile F1 offspring of both sexes that can reproduce without backcrossing with a parent species (even if there are observed asymmetries in gene flow and variation in male fertility among the hybrids) (electronic supplementary material, table S1). category 2 is defined by pairs of species that can produce viable F1 offspring, but follow Haldane's Rule, and thus only female F1s can reproduce by backcrossing with a parent species. category 2 also includes species pairs whose hybrids are infertile (electronic supplementary material, table S1). We determined the categorical assignment of each species pair (electronic supplementary material, table S1 and figure S1) by following a decision tree (electronic supplementary material, figure S2) based upon empirical evidence derived from experimental studies of F1 hybrid fertility. We confidently placed seven species pairs into category 1, and six others into category 2. Many additional live hybrid offspring have been reported in the literature than are included in figure 1 or in the electronic supplementary material, figure S1. We identified 17 species pairs known to produce viable offspring, but for which there was insufficient evidence to confidently assign them into either category (electronic supplementary material, table S2 and figure S3). The framework and threshold values depicted in figure 1 allow us to predict the fertility of these offspring given the definitions described above, and their placement into categories 1 or 2. These pairs are listed in the electronic supplementary material, table S2 and their relative positions are depicted in the electronic supplementary material, figure S3. Figure 1. A depiction of the correlation between CYTB divergence between mammalian species pairs and the relative fertility of their hybrid offspring. In column (a), the green circles represent species capable of producing fully fertile F1 offspring which can reproduce independently of their parent species (category 1). Brown circles represent species pairs that follow Haldane's Rule and require backcrossing of a female F1 with a parent species, or both sexes are sterile (category 2). The single grey dot represents the distance between mountain hares and European rabbits that, despite numerous attempts, failed to produce any offspring. The green and brown shaded regions represent the range of divergence values of the two categories. Column (b) depicts the divergence between three wild felid species and domestic cats, as well as the minimum number of generations of backcrosses with domestic cats before full fertility of the hybrid is restored. The white circles in column (c) depict the divergence between three ancient hominins and anatomically modern humans (AMH), as well as the distances between AMH and chimpanzees and bonobos (in category 2). The asterisks represent those pairs that include modern samples of AMH. The lack of an asterisk signifies that only sequences derived from archaeological AMH were used to compute the divergence values. Details regarding the specific species pairs are listed in the electronic supplementary material, figure S1 and table S1. (Online version in colour.)
We then obtained published sequence data across all species (electronic supplementary material, table S3) from both the cytochrome b gene (CYTB) (n = 1795) and complete mitochondrial genomes (n = 30) (excluding the control region) from multiple individuals per species. By matching the phylogenies derived from the alignments to available nuclear species trees, and by only including sequences that fell into reciprocally monophyletic clades, we ensured that the selected mitochondrial sequences for each species were neither mislabelled, nor nuclear copies of mitochondrial genes (NUMTs), nor derived from hybrid populations. Using the sequence alignments, we then calculated average pairwise genetic distances between each species pair using both the number of raw differences, and differences scaled by several different nucleotide substitution models. In order to test the correlation between the mitochondrial proxy and estimates obtained from nuclear loci, we calculated genetic distances for four genes (cholinergic receptor nicotinic alpha (CHRNA1), growth hormone receptor (GHR), zinc finger X-chromosomal protein (ZFX), and zinc finger Y-chromosomal protein (ZFY)) that were available for 10 of the primate species pairs [23] (electronic supplementary material, figure S4). Plotting the calculated distance values using CYTB revealed a limited overlap associated with the two categories (figure 1; electronic supplementary material, figure S1), and a Student's t-test (two tailed) showed significantly lower genetic divergence values of species pairs in category 1 relative to those in category 2 (p < 0.003). The category 1 pair with the greatest divergence value was a pair of guinea pig species that were 8.0% divergent, and the category 2 pair with the lowest divergence was a pair of vole species that were 7.2% divergent. Several pairs of species fall within this 0.8% overlapping region suggesting that this level of CYTB distance is the zone where some F1 mammalian offspring begin to require a backcross to generate an F2. The existence of a genetic distance threshold separating the two categories also held true for the complete mitogenomes (electronic supplementary material, figure S4). Towards the upper end of distance values between species pairs, both the male and female hybrid offspring of domestic pig (Sus domesticus) × babirusa (Babyrousa celebensis) (12.9%) were shown to be infertile [24]. In addition, controlled, exhaustive efforts failed to produce any viable hybrids between mountain hares and European rabbits (17.3%) [25] (figure 1). The large distance values between these species pairs corroborates previous studies showing that along the continuum of speciation, infertility in both sexes evolves prior to inviability [26–28]. Importantly, the two categories of fertility defined here are not strictly linked with gene flow. For instance, though both male and female category 1 hybrid offspring can reproduce without requiring a backcross with a parent species, gene flow asymmetries have been demonstrated in virtually all of these species pairs including house mice (Mus musculus musculus × Mus musculus domesticus) (2.3%) [29], and between brown bears (Ursus arctos) and polar bears (Ursus maritimus) (2.4%) [2] (electronic supplementary material, table S1). Gene flow has also been demonstrated between category 2 species (including M. musculus and Mus spretus [30]) (8.6%). Because both fertility and the potential for gene flow vary along a continuum, it is striking that the divergence values associated with the two fertility categories defined here do not overlap more substantially. The lack of available nuclear sequences (relative to mitochondria) reduced our ability to test whether nuclear genes generally produced the same pattern as the mitochondria across all our species pairs. Despite this limitation, we were able to identify four nuclear loci: ZFY, ZFX, GHR, and CHRNA1, that have been sequenced in 10 primate pairs (electronic supplementary material, figure S4) known to produce viable hybrid offspring [23]. We generated pairwise distances for each of these genes using the same method employed in the mitochondrial distance calculation. We then assigned each species pair to category 1 or category 2 based upon their CYTB divergence values within the original framework. In each case, though the order of the taxa based upon pairwise divergence values varied relative to the pattern generated using CYTB (owing to significantly fewer variable sites and thus smaller divergence values in nuclear loci), for the two most variable nuclear loci, ZFX and GHR, there was no overlap in divergence values between the two categories, consistent with the mitochondrial assessment (electronic supplementary material, figure S4). The other two loci, ZFY and CHRNA1, possessed very limited interspecific nucleotide variability, but generally followed the same overall pattern. In order to further substantiate both this correlation and the robustness of CYTB divergence as a proxy for hybrid sterility, we tested the use of this system for predicting fertility in a well-known hybrid system. Cat breeders have crossed domestic cats (Felis catus) with several wild felids, including the jungle cat (Felis chaus), leopard cat (Prionailurus bengalensis) and serval (Leptailurus serval) [31], to create three exotic cat breeds: Chausies, Bengals and Savannahs, respectively. In all cases, the F1 male hybrids are sterile. To regain fertility while maintaining some wild felid characteristics, breeders must backcross the F1 female offspring with male domestic cats to establish a breeding population of pets [31]. Given that multiple generations of unidirectional backcrossing was required for all three crosses to generate a fertile population, our proxy would firstly predict that the CYTB distances between all three pairs should be close to or greater than approximately 7.2%, and that they should all fall into the range encompassed by category 2. Secondly, the pairs with larger genetic distance values should require a greater number of backcrosses with domestic cats (halving the wild cat ancestry with each subsequent generation) before fertility is restored in hybrid males and a breeding pet population is established. Both of these predictions are borne out by the data (figure 1; electronic supplementary material, table S3). All three pairs show CYTB distances greater than or equal to 7.5% and the increasing molecular distances between the pairs correlate with an increase in the number of required backcross generations to regain fertility. Specifically, distances between domestic cats and jungle cats, leopard cats and servals (7.5%, 10.9% and 11.3%, respectively) are consistent with both the observed minimum (2, 3 and 4, respectively) and average (3, 4 and 5, respectively) number of backcrosses with domestic cats required for hybrid males to acquire fertility [31]. These results are also consistent with an early hybrid experiment using guinea pigs in which hybrids between Cavia fulgida and Cavia porcellus (8.0% CYTB distance) were able to regain male fertility after three generations of backcrossing [32] (electronic supplementary material, table S1). Accidental hybrids in zoos also confirm the predictive power of this proxy. In 2006, the Copenhagen Zoo placed a domestic sow (S. domesticus) in a pen with a male babirusa (B. celebensis) with the expectation that the two species were sufficiently evolutionarily divergent that they would be incapable of producing offspring. Months later, however, five piglets were born and though two died from maternally induced trauma, the other three (two males and one female), all survived and were shown to be infertile [24] (electronic supplementary material, table S1 and figure S1). Historically, hybrid offspring between distantly related species have accidentally been produced in zoos (electronic supplementary material, table S2), though the relative fertility of the F1s was rarely established. In this case, the CYTB distance between the two species (12.9%) is not much greater than the value between rhesus macaques and hamadryas baboons (12.5%) which were able to produce a live (infertile) offspring [33], thus suggesting that live offspring between these suids was possible. The initial discovery of Neanderthals led some anthropologists, as early as the turn of the twentieth century, to speculate that anatomically modern humans (AMH) and their closest extinct relatives were capable of producing hybrid offspring [34]. The absence of Neanderthal mitochondrial genomes in the extant human population, however, led some to suggest that AMH and Neanderthals did not hybridize [35–37]. More recent analyses of whole ancient genome sequences have demonstrated that, in fact, archaic hominins, including Neanderthals and Denisovans, did produce hybrid offspring, not only with AMH [38–40] but also with each other [41]. The generation of these ancient genomes has also allowed for an assessment of the role that incompatibility may have played in the selection for and against hybrid introgression in modern humans [42]. The genomic confirmation of the existence of hominin hybrids supported the conclusions of two studies [43,44] that used a qualitative correlation between the divergence times between species pairs and the fertility of their hybrid offspring to suggest that, given their relatively recent temporal divergence, AMH and Neanderthals could have retained the ability to produce fertile offspring of both sexes. We quantitatively assessed the relative fertility of hybrids between pairs of modern and ancient hominin lineages using the proxy established in this study. To do so, we calculated the average pairwise distance in CYTB sequences between AMH and three extinct hominin lineages: Neanderthals, Denisovans and the ancient population from the Sima de los Huesos cave in Spain [24,45]. To avoid overestimating the genetic distances resulting from the comparison of modern and extinct populations, we generated values using the CYTB sequences derived solely from ancient AMH found in archaeological contexts (electronic supplementary material, table S3). The distance values for all of the pairings of three Homo groups (Sima de los Huesos, Neanderthals and AMH) occupy the bottom of the category 1 range. The distance values for Neanderthals and modern and ancient AMH specifically (1.6%) fall below all the mammalian pairs in this study including polar bears and brown bears (2.4%), and between subspecific crosses of M. musculus (2.3%) (figure 1; electronic supplementary material, figure S1 and table S1). When placed within this context, our data predict that ancient hominin lineages were probably not sufficiently divergent from each other to expect a significant biological impediment to the generation of fertile offspring. This is consistent with the ancient genomic evidence, which has shown not only that archaic populations interbred with AMH on at least four occasions [46], but also that introgression took place in both directions [47]. In addition, the distance values of Denisovan-Neanderthal and Denisovan-AMH are the largest of the Homo pairings, and are consistent with the suggestion that Denisovans possessed a mitochondrial lineage that may have been acquired through introgression with another, as yet unknown source population [48]. We also assessed hybrid sterility between more distantly diverged hominin lineages. Specifically, we calculated divergence values between humans and our two closest living relatives: chimpanzees (Pan troglodytes) and bonobos (Pan paniscus). Female chimpanzees inseminated with human sperm during a Soviet experiment in the 1920s failed to produce any offspring, and the reverse experiment did not progress beyond the planning stage [49]. Recent molecular clock assessments have suggested that AMH and chimpanzees diverged approximately 5–6 Ma [50], well beyond both the 2 Myr threshold suggested by other studies as the upper limit to hybrid fertility [43,44], and the average time to speciation [51]. Our analysis places the distance values between AMH and chimpanzees (11.0%), and AMH and bonobos (10.8%) within category 2, suggesting that even if hybrids could be produced, they would probably follow Haldane's Rule (figure 1; electronic supplementary material, figure S1). The correlation demonstrated here between CYTB divergence (as well as genetic divergence in general) and relative hybrid sterility suggests that distance values can be used as a proxy to accurately and rapidly predict the relative sterility of hybrids that result from matings between pairs of mammalian species. More specifically, our results show that the F1 offspring of some mammalian species pairs with greater than 7.2% CYTB distance have lost the ability to produce F2s, and beyond 8.0%, all pairs of species in our dataset require a backcross to a parent species to produce fertile offspring. In addition, our results demonstrate that once a single backcross with a parent species is required to restore fertility in hybrid males, the number of additional necessary backcrosses increases with greater CYTB distances between the parent species. Our emphasis here is on mitochondrial DNA, and though recent studies have proposed that speciation may be mediated by mitonuclear interactions [52,53], our results should not be misinterpreted as a claim that CYTB plays a causative role in hybrid sterility. Nor can the use of genetic distance values as a proxy be perfectly predictive. For example, under the Dobzhansky–Müller model, incompatibility can arise from as few as two mutations in isolated populations irrespective of time since divergence. This means that it would be possible for closely related populations to be incapable of generating fertile hybrids [52,54], though no such examples have been described. The value of any proxy is determined by both its predictive power, and the ease of generating the proxy data. Publicly available mitochondrial DNA sequences from thousands of mammalian taxa already exist and calculating pairwise distance values is inexpensive, simple and fast. As a result, mitogenomic distances have substantial value as a means to predict the potential for any two mammalian species to produce fertile offspring, and the relative degree of sterility in one or both sexes. As whole genomes become available from the same set of species, this analysis can be extended to determine which regions of the nuclear genome may also be more or less predictive. The discovery of additional extinct hominin populations that survived into the last 250 000 years, including Homo floresiensis [55] and Homo naledi [56], has raised interest in understanding the limits to fertility and hybridization between extinct and extant Homo spp. [57]. If and when mitochondrial genomes from these samples can be obtained, the approach described here may provide an answer, even if nuclear genomic data are not obtainable. Lastly, establishing which species pairs violate the predictions of the framework will identify unique systems that may lead to a better understanding of the process of reproductive isolation, and the biological mechanisms responsible for hybrid sterility. In order to ascertain if there was a correlation between genetic divergence and the fertility of hybrid offspring between species, we first collected published examples of species pairs that were capable of producing live offspring. We then split the hybrid pairings into two categories. category 1 consisted of seven species pairs that are capable of producing fertile F1 offspring of both sexes, and for which we were able to obtain evidence of captive breeding experiments showing that the F1s could mate to produce F2s. The evidence and rationale for placing each of these pairs into category 1 is listed in the electronic supplementary material, table S1 and the decision tree we used to determine the categorization is shown in the electronic supplementary material, figure S2. The hybrid offspring of all of six pairs of species in category 2 are either completely infertile, or require one or more generations of female hybrid backcrosses with the male of a parent species to produce fertile offspring. For these pairs, we obtained evidence demonstrating no successful F2s from F1 hybrid pairings, an inability to produce offspring other than by backcrossing to a parent species, or other biological measurements (including histological assessments of the testes from the hybrid males) that demonstrated complete infertility (electronic supplementary material, table S1 and figure S5). Both CYTB sequences and full mitogenomes (excluding the control region) of multiple individuals of each species were collected from Genbank (electronic supplementary material, table S4) and aligned using Clustal Omega v. 1.2.4 [58]. In order to ensure that none of the sequences were either mislabelled, or were NUMTs, we constructed neighbour-joining trees using Geneious v. 6.1.8 [59] and removed all individuals that did not fall into monophyletic clades consisting of individuals from each species. We first used pModelTest v. 1.04 [60] to determine the best model for the alignment of each set of sequences for both species. We then calculated pairwise distances between each species pair using RAXML v. 8 [61] and FastTree v. 2.1 [62]. We also generated raw distance values using the Hamming distance method which sums the number of base pair differences (ignoring transition or transversion status) and divides that number by the sequence length. The distances were generated from the CYTB and nuclear gene alignments for each set of species pairings in fasta file format using a Python v. 2.7 wrapper to automate the terminal based programs RAXML, FastTree and pModelTest. A custom Python 3.6 program was written to calculate Hamming distances of sequences making use of the distance v. 0.1.3 [63] and Biopython v. 1.66 [64] modules. Gaps in the aligned sequences were treated as missing data. The mean distance and standard errors for each pairwise comparison were calculated using the bootstrapping method on the assumption that the sets of pairwise distances between related species would not be normally distributed. Each pairwise comparison group containing Hamming distances was randomly resampled into sets of equal sample size and processed using a helper function in the custom software which made use of the bootstrapped v. 0.0.2, NumPy v. 1.10.1 [65] and SciPy v. 0.16.0 [66] Python modules. The source code of this script is available at https://github.com/BeebBenjamin/MrHamming. The CYTB distances were compared with those generated in MEGA X for GNU/Linux [67] (which uses a slightly different method for treating missing bases) using the ‘compute between group mean distance’ method with the following settings:
The statistical significance of observed differences in CYTB divergence between categories 1 and 2 was tested using the Student's t-test (p = 0.002942) implemented in the R software package [68]. The suitability of a parametric test was determined using the Shapiro–Wilk normality test (p = 0.6238). Data has been uploaded as part of the electronic supplementary material. R.A. and H.R. compiled and analysed data and wrote the paper. B.W.D., C.K., R.B., A.L., L.L., J.H., O.L., M.W., A.C.K. and L.F. wrote the paper. E.I.-P. analysed data and wrote the paper. W.J.M. generated data and wrote the paper. G.L. conceived of the study and wrote the paper. We declare we have no competing interests. G.L. was supported by the European Research Council (grant no. ERC-2013-StG 337574-UNDEAD) and the Natural Environment Research Council (grant nos. NE/H005269/1 and NE/K005243/1). W.J.M. was supported by the National Science Foundation (grant no. DEB-1753760). We thank Simon Ho, Linda Maxson, Julie Wilson, Andrew Millard, John Hawks, Tom Higham, Kelly Harris, Christian Capelli, Shyam Gopalakirshnan, Montgomery Slatkin, Joshua Schraiber, Sam Turvey, Janet Kelso, Al Roca and Murray Cox for advice and discussion. We also thank the Zoological Society of London and the Bartlett Society for their assistance. Footnotes†These authors contributed equally to this study. Electronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.4979948. References
Page 19Start a war or negotiate peace? Invest in more capital stocks or sell shares? The fate of states, companies and organizations are shaped by their decisions. It is then surprising that only a minority of individuals are involved in the decision-making process. From companies to political parties, organizations tend to follow an ‘iron law of oligarchy’, in which larger and more productive groups switch to hierarchy where a few individuals possess most of the political power, resources and influence [1]. This transition is best illustrated by the deep overhaul of human societies initiated by the advent of agriculture 12 500 years ago. In a relatively short period of time, most human groups switched from non-hereditary and facultative forms of leadership [2,3], to hierarchical societies with one or few permanent leaders [4]. Despite this transition being well known, it is still hard to explain why human groups follow this general trend. The independent transitions to hierarchy and its pervasive presence suggest that the emergence of hierarchy is (at least partly) the result of natural selection [5]. But the emergence of hierarchy appears as a Darwinian paradox because leaders often enjoy preferential access to resources [4] and mating partners [6]. Why would any individual rationally accept a position of being a follower if the position of leader is more beneficial? One could argue that followers do not have a choice because leaders impose their dominance using coercive means. Humans have inherited traits and preferences towards hierarchy from their primate ancestors, who were organized in dominance hierarchies where individuals physically compete for rank, resources and partners [2]; but adaptations such as the capacity to form large coalitions and the development of throwing weapons led early human groups to reverse this hierarchy [7,8]. In pre-Neolithic tribes, coalitions of followers imposed strong dominance on uprising leaders and successfully avoided coercive leaders for hundreds of thousand of years [2]. Therefore, coercive theories fail to explain the emergence of hierarchy in the first place when any advanced form of coercion, e.g. armies, taxes or propaganda, was absent. An alternative theory from political sciences proposes that in the absence of financial or military power, leaders have established their dominance by first accumulating political power, i.e. influence over collective decisions [1]. The ‘iron law of oligarchy’ states that political leaders inevitably arise as a group expands, in order to deal with the complexity of coordination. But, this allows these leaders to bias collective decisions in their favour, e.g. distribution of resources or command of a military. This theory fits well with evidence of leaders playing a prevalent role in group organization [9]. As can be seen in small scale societies [10] or in a modern share-holder meeting [11], leaders reduce the cost of organization by assigning roles to individuals, settling arguments between decision makers, and helping to decide the future course of action. However, this theory suggests that the group benefit would be enough to overcome individual selection driving everyone to become a leader rather than a follower [12]. There is clearly a conflict between individual and group interests, which makes this condition non trivial [13]. The ‘iron law of oligarchy’ proposes that a key element is scalar stress, which describes the fact that the difficulty for a group to coordinate increases with group size [1]. This relation appears in (i) psychology experiments of collective decision-making, in which larger groups reach a lesser degree of agreement [14] or take worse decisions [15], and (ii) indirectly in anthropological data showing a strong correlation between group size and probability of group fission [16], or group size and the number of political units [17]. On one side of the range, small-whale hunters of Mackenzie Inuits have one single coach to coordinate group hunting [18]. On the other extreme, complex states or companies have dozens of politicians and managers who are fully dedicated to the task of organizing. Previous work has shown that scalar stress can drive the evolution of institutionalized hierarchy [19], where a leader is appointed by a centralized process. However, rather than being ascribed, hierarchy is likely to initially emerge from the evolution of intrinsic physical and psychological traits of individuals, e.g. height [20], talkativeness and charisma [21]. Emblematic examples of such informal hierarchy are the ‘big man’ societies observed in Melanesia, in which leaders are defined by their persuasion skills rather than by an ascribed position [22]. However explaining the evolution of ‘informal’ hierarchy without supporting institutions poses an important challenge: can the group benefit of hierarchy overcome the selection pressure pushing everyone to be a leader? The lack of a mechanistic model describing the effect of hierarchy on collective decision-making has limited investigation of scalar stress as a possible solution. The iron law of oligarchy [1] and behavioural experiments [23] suggest that the benefit of hierarchy on group coordination lies in its effect on the time a group spends to reach consensus and take a collective decision. Consensus decision-making is an efficient method for a group to coordinate, in particular to tackle tasks where learning the optimal strategy by trial and error is too costly. Examples of consensus decision-making include tribe gatherings to discuss the next camp location, councils of war to decide upcoming battle strategies, or parliamentary debates on new laws. Yet, the time spent to reach consensus (consensus time in short) is costly because individuals dedicate time to organization instead of carrying out the actual task, and because time itself can carry a cost, e.g. resources get depleted. Thus, we explicitly model the consensus decision-making and the effect of hierarchy on this process. We describe group social organization as a distribution of individuals’ influence, i.e. their capacity to modify another individual’s opinion towards their own. The scale from acephalous to highly hierarchical groups is represented by an equal to strongly positively-skewed distribution of influence. We use the term leaders and followers to describe individuals with high and low influence, respectively. This definition of hierarchy does not include the degree of inequality in resources. We allow the correlation between hierarchy and degree of inequality in resources to emerge from the model, as influential individuals can bias the distribution of resources to their advantage. The emergence of hierarchy is represented by the evolution of individual behaviours towards a minority of leaders and a majority of followers. We combine social and evolutionary dynamics to investigate the development of hierarchy and build a mechanistic formalization of the ‘iron law of oligarchy’. Using this model, we aim to answer the following question: does hierarchy limit the effect of scalar stress, and if yes, could it drive the evolution of leader and follower behaviours even if it creates inequality in resources? To do so, we consider that in absence of advanced institutions such as voting systems, collective decision-making is a sequence of communications, as observed in human groups faced with coordination problems in laboratory experiments [24] or in the real world [2,25]. Thus, we mathematically describe collective decision making by an opinion formation model, which consists of a sequence of discussions between individuals until a global consensus is reached [26]. We investigate how social organization affects scalar stress, where scalar stress is defined as the relationship between the time spent to reach consensus and group size. It has been hypothesized that scalar stress is enough to drive the evolution of stable hierarchy [27]. We test this hypothesis in the second half of the paper by explicitly integrating the consensus decision-making into an evolutionary model. We consider a population structured into patches, where individuals on a patch organize together to produce a collective good. The consensus time determines their cost of organization, and the influence of an individual on the final decision determines that individual’s share of the collective good. We developed an opinion formation model based on previous work [28], which simulates a sequence of discussions between N individuals until consensus is reached. The outline of the model is represented in the electronic supplementary material, figure S1. Individuals are represented by an opinion x. The opinion x describes a generic opinion of an individual on how to realize a collective task, e.g. the next raid target, the plan of an irrigation system or the value of a law. Individuals are also described by a value of influence, α. The influence, α, is defined as the capacity of one individual to influence group decision and affects: (i) the capacity of one individual to modify the opinion of another individual towards their own opinion, (ii) the reluctance of an individual to change their opinion, and (iii) the probability that an individual talks to other individuals. These three traits, i.e. persuasiveness, stubbornness and talkativeness, are highly correlated in leaders' personalities [21] and are the key factors in explaining how leaders affect consensus decision-making [28]. Both the opinion x and the influence α are continuous variables defined on [0,1]. To facilitate the analysis of the opinion formation model, we divide individuals into two profiles: leader L, and follower F. Each profile has a fixed value of influence α such that αL > αF. At the beginning of the opinion formation model, the values of opinion x are sampled from the uniform distribution between [0, 1]. Each time step is defined by one discussion event during which one speaker talks to multiple listeners. The probability P of an individual i to be chosen as a speaker is an increasing function of its α value as follows: Pi(t)=(αi(t))k∑n=1N(αn(t))k.2.1 The exponent k defines how much the difference in influence is translated into a difference in the probability to talk. In the simulations we chose k = 4 so that in a group of 1000 individuals with the most extreme hierarchy, the probability that a leader is chosen as a speaker is very high (close to 90%).The speaker talks with Nl listeners, with listeners being a subset of the total group. We assume that the number of listeners is limited because of time constraints. The listeners are randomly sampled from the other individuals in the group. We assume that every individual can be chosen as a listener, i.e. the social network is a complete network. During a discussion event, a listener v updates its opinion to a value x′v following the equation below, where v represents the listener and u the speaker: xv′=xv+(αu−αv) (xu−xv).2.2 We assume that the position of speaker gives a slight influential advantage over the listeners. Therefore, the minimum difference of influence αu − αv is set to a positive low value, here 0.01. This assumption is necessary to avoid a systematic convergence of individual opinions towards those of the individual with the highest α, a phenomenon not observed in real life. The individuals repeat the previous step until the following condition is true:σx<xθ.2.3 Consensus is reached when the standard deviation of the opinions σx is inferior to the threshold xθ. The number of discussion events that occurred to reach consensus is called the consensus time, t*. The final decision reached, x*, is the mean of the opinions at consensus across individuals. Using this model, we aim to answer the following question: does hierarchy limit the effect of scalar stress?Figure 1 shows that the presence of few leaders (i) reduces the consensus time, and (ii) reduces scalar stress, which is shown by the gradient of consensus time with respect to group size being smaller. In other words, hierarchy facilitates organization by reducing the intensity of scalar stress. Importantly, the differential quality of information that leaders might posses, and which might lead to a group with hierarchy making better decisions, is not required to get this result. Figure 1. (a) Consensus time as a function of the size of the group for three different types of social organization: (i) 0 leaders, (ii) 1 leader and (iii) 10 leaders. (b) Scalar stress measured by the linear regression coefficient (slope) of time to reach consensus on group size as a function of number of leaders. The ribbons represent the standard deviation to highlight the high variance in the consensus time when multiple leaders are present. One hundred independent replicates have been realized for each group size and social organization. The parameters used are Nl = 30, αL = 0.75, αF = 0.25 and xθ = 0.05. (Online version in colour.)
The electronic supplementary material, figure S2 shows that hierarchy reduces scalar stress and that this result is consistent across different leader and follower profiles. It shows that hierarchy with a single leader has the lowest scalar stress when the difference in influence between leaders and followers is high. Conversely, hierarchy with multiple leaders has the lowest scalar stress when the difference in influence between leaders and followers is low. This is because multiple influential leaders (i) can increase divergence by convincing followers towards extreme opinions, (ii) can convince followers from other leaders, and (iii) are slower to themselves be convinced. The electronic supplementary material, figure S3 demonstrates that the scalar stress is strongly dependent of the number of listeners Nl. It shows that a lower number of listeners Nl results in an increase in consensus time, in particular for acephalous groups and multiple leaders hierarchy. This is because a lower number of listeners slows down the convergence of opinions. In the presence of multiple leaders, it also reduces the probability that one leader convinces the majority of the group and thus, favours the formation of clusters of different opinions built around one stubborn leader. In conclusion, our model demonstrates scalar stress, the benefit of hierarchy in reducing scalar stress, and how this benefit is amplified by limited interactions as is the case with a low number of listeners Nl. We have shown that the presence of a minority of influential individuals (leaders) and a majority of influenceable individuals (followers) reduces the cost of organization and scalar stress. Can this organizational advantage be enough to drive the evolution of individuals towards leader and follower behaviours even if it creates inequality in resources? We answer this using an evolutionary model where we describe social organization as a distribution of influence, and use opinion formation to link this distribution back to the cost of organization. The life cycle of the evolutionary model is represented in the electronic supplementary material, figure S1. We now let the trait α carried by individuals evolve. The trait α is transmitted vertically from parent to offspring, e.g. by social learning, as is common in both hunter–gatherer groups [29] and modern societies [30]. Unlike the model above, the possible values of influence α can take any value in the range [0,1]. We use the terms leader and follower to designate individuals with high and low influence, respectively. The trait α mutates following a mutation rate of μ. As α is likely to be at least partly cultural, we assume a mutation rate higher than for a classical genetic trait. When a mutation occurs, a random value is sampled from a truncated Gaussian distribution centred on the current value of the trait, with variance σm2. We study the evolution of α using a standard island model with a population of individuals that is subdivided into a finite number of patches, Np [31]. We consider that group size can vary and thus, groups can compete by differential migration. The life cycle of individuals consists of discrete and non-overlapping generations, where in each generation the following occur: (i) consensus decision-making about how to perform a task; (ii) performance of the collective task; (iii) distribution of resources obtained from the task; (iv) reproduction; and (v) migration. The first three steps determinate the reproductive success of an individual, which we denote by its fitness w. The fitness of an individual is translated into a number of offspring, which is drawn from a Poisson distribution centred on its fitness w. After reproduction, offspring individuals migrate with a probability equal to a fixed migration rate m. Migrating individuals enter a patch chosen at random from the population (excluding their natal patch). More formally, the fitness w of individual i on patch j at time t is described by the following equation: wij(t)=ra1+(Nj(t)/K)+rbij(t),3.1 where Nj(t) is the total number of individuals on patch j. The fitness of an individual is the sum of (i) an intrinsic growth rate ra, limited by the carrying capacity K, and (ii) an additional growth rate resulting from the extra resources produced by the collective task, rbij(t). The additional growth rate rbij(t) is not limited by the carrying capacity, but competition between individuals is taken into account during the distribution of collective resources. The additional growth rate rbij(t) is calculated as follows:rbij(t)=βr(1−e−γr(Bj(t)pij(t))).3.2 The term rbij(t) is calculated from a logistic function described by γr and βr, respectively the steepness and the maximum of the increase in growth rate induced by the additional resources. The additional resources are given by the total amount of benefit, Bj(t), multiplied by the share the individual receives, pij(t). The increase of the growth rate follows a logistic relation because of the inevitable presence of other limiting factors, e.g. space or other resources.To produce the additional resources Bj(t), individuals first undergo a consensus decision-making process on their patch, as defined in the previous section (see equations (2.1)–(2.3)). The consensus time determines the cost of organization (equation (3.3)), and the outcome of the consensus decision-making determines the share of individuals (equation (3.4)) as explained in the following paragraphs. We do not consider that the decisions taken affect the success of the group except for the distribution of resources. We investigate the emergence of leaders defined as influential individuals, as it was done in psychology experiments [32], and as observed in psychological profiles of leaders [21]. In this case where leaders are not better at taking decisions, integrating the effects of the collective decision would only result in more noise and not qualitatively change our results. Exploring the emergence of leaders as more informed individuals can be done in further work but it is not our focus here. After consensus is reached, all individuals on a patch take part in the collective task which produces an amount of extra resource Bj(t): Bj(t)=Bj(t−1)S+βb1+e−γb(Nj(t)−bmid)−Cttj∗.3.3 The benefit is calculated from a sigmoid function described by βb, bmid and γb, respectively the maximum, the group size at the sigmoid’s midpoint, and the steepness of the increase in the benefit induced by additional participants. We make the assumption of initial increasing returns to scale, where additional participants increase the benefit superlinearly [33]; but as is standard in microeconomic theory, we also make the conservative assumption that the benefit of the collective task eventually has diminishing marginal returns which overcomes the increasing returns to scale because of other limiting factors, e.g. land available or level of technology [33]. To capture the transmissibility of resources [34], we assume that a proportion S of the benefit is also present in next generation. The extra resources are discounted by a cost of organization proportional to the consensus time t*.This cost is modulated by a parameter Ct, which describes the time constraints on group decision-making. The parameter Ct depends on the pressure of time on the task, for instance, the speed of depletion of resources or the need to build defences before an enemy arrives. To avoid studying the effect of social strategy in the collective task, which has already been extensively studied in the evolution of cooperation literature [35], we consider that all the individuals on a patch are willing to participate in the collective task once a decision is reached. The collective task simulates the numerous cooperative tasks realized during the lifetime of an individual, e.g. hunting large game or constructing an irrigation system. The resources obtained from the collective task are distributed among all individuals on the patch. We want to test if hierarchy can emerge even if leaders receive a higher share of the collective resources, which selects against individuals becoming followers. However, leaders are not clearly designated in informal hierarchy. We assume that in the absence of coercive means, individuals can only increase their share by biasing the collective decision towards their own interests and thus, the share of an individual pij(t) is a function of its realized influence αr such that: pij(t)=1+dαr(ij) (t)∑i=1Nj(1+dαr(ij) (t)).3.4 The asymmetry of the distribution of the resources is modulated by a parameter d, which represents the level of ecological inequality. For d = 0, a patch is totally egalitarian and the influence of an individual does not affect the share of that individual. Such a scenario is close to the society of pre-Neolithic hunter–gatherers [2]. It is assumed for simplicity that d is the same for all patches, and is determined for example by the state of technology, such as food storage and military technologies. Nevertheless, different patches can have more or less despotic distributions of resources owing to different distributions of αr values. The realized influence of an individual αr(ij) is calculated from the difference between an individual’s initial opinion and the final decision, and measures how much the final decision is close to the individual’s interest:αr(ij)=1−|xij(t=0)−xj∗|.3.5 We use this model to answer the following question: can the organizational benefit of hierarchy in the presence of scalar stress lead to the evolution of leader and follower behaviours even if it creates inequality in resources? Because of the nonlinearities of the model, we analyse it using replicated individual-based simulations. We define hierarchy as a positively skewed distribution of influence α. The skewness is measured by the Pearson’s moment coefficient of skewness. We focus on the effect of the following parameters: (i) the level of ecological inequality d, and (ii) and the number of listeners Nl, which affects the intensity of scalar stress (electronic supplementary material, figure S3). In the electronic supplementary material, we also explore the effect of (i) the time constraints on group decision-making Ct, (ii) the migration rate m, which affects the population structure, and (iii) the absence of transmission of resources from one generation to another (S = 0). The default values of these parameters, unless otherwise specified, are for the level of inequality d = 1, the number of listeners Nl = 30, the time constraints on group decision-making Ct = 2, the migration rate m = 0.05, and the fraction of resources transmitted to next the generation S = 0.9. The default values for the remaining demographic and ecological parameters are for the number of patches Np = 50, the initial number of individuals on each patch Nj(0) = 50, the carrying capacity K = 50, the intrinsic growth rate ra = 2, βb = 10 000, γb = 0.005, bmid = 500, βr = 3 and γr = 0.025. These values are chosen in order to allow the transition between tribe size (50–100 individuals) to small chiefdom size (500 individuals) [4], and so that additional resources lead to a clearly increased fitness. The remaining default parameter values are for the consensus threshold xθ = 0.05, and for the mutation process, where μm = 0.01 and σm2=0.01. The results presented are the average across patches when the result is a function of generations, and the average across patches, generations and simulations when the result is a function of a parameter. The error bars represent the standard error from the mean across replicates. The simulations are run for 10 000 generations and the first 5000 generations are ignored to limit the effects of initial conditions. Figure 2 presents the evolution of the distribution of influence and group size as a function of generations for a single run. The results show that despite the wide range of possible distributions of influence, individuals evolve towards hierarchy, i.e. a minority of leaders with high influence and a majority of followers with low influence. In the meantime, the population grows to a large group size. Within a patch, hierarchy also evolves but the proportion of leaders and followers vary. The result is stable across replicates and in the long term as shown by figure 3a. At the start of the simulation, groups have a skewness close to 0 and a small group size because the values of influence are randomly initiated. Figure 3a demonstrates that skewness increases with time and remains at a positive value along generations. The positive skewness reflects a majority of individuals with low influence—followers—and a minority of individuals with high influence—leaders. This result is also present in absence of intergenerational transmission of resources (S = 0) as seen in the electronic supplementary material, figure S4. Overall, these results show that hierarchy can emerge from the evolution of individual behaviour and thus, hierarchy provides a clear evolutionary advantage. Figure 2. Evolution of the distribution of influence α, and evolution of average group size as a function of generations for the whole population (a) and three different patches (b). The plot represents results for a single run. A column at a given generation is composed of sections with each section showing the proportion (size of the section) of individuals with a given influence (colour of the section). Note that there is a small proportion of individuals with high influence at equilibrium. This proportion is low and thus, hard to discern but it is revealed by a stripe of orange and white colour at the bottom. (Online version in colour.)
Figure 3. Evolution of (a) the average skewness of distribution of influence α, (b) average group size, and (c) average consensus time t* (red) and average additional resources produced b (blue), as a function of generations. Hierarchy is represented by a positive skewness. The values presented are the average across 32 replicates. (Online version in colour.)
The benefit of an efficient hierarchical organization is shown in figure 3c. It shows that over generations, the consensus time and the total amount of resources both increase. This is because group size increases and leads to more resources being produced owing to increasing returns to scale, but also a greater difficulty to organize. However, it can be observed in figure 3c that the increase in consensus time stabilizes before the end of the increase in extra resources. This is because individuals have evolved towards hierarchy and can maintain a low cost of organization even as the group size and the production of resources continue to increase. The benefit of hierarchy depends on the time constraints Ct, which translates the consensus time into an opportunity cost of organization. The electronic supplementary material, figure S5 shows that the level of hierarchy is proportional to the time constraints. For a low level of time constraints, the benefit of hierarchy has a negligible effect on organization and group production and thus, hierarchy does not evolve. For tasks with strong time constraints, e.g. warfare, the benefit of hierarchy is amplified and a strong hierarchy evolves. Hierarchy evolves because it reduces the cost of organization and thus provides the creation of a surplus in group production. This surplus is distributed among the individuals and increases the number of offspring they produce. This results in hierarchical groups growing larger and exporting a larger number of migrants than groups without hierarchy. Most of these migrants are followers because most of the population within a hierarchy are followers. Ultimately, these migrants spread hierarchical organization to other groups and at the population level creates a stable skewed distribution of influence. Importantly, hierarchy evolves even when the emergence of hierarchy creates inequality in resources. Hierarchy creates inequality in resources because leaders will more often bring the group decision close to their preferences and thus receive a higher share of the resources produced, and hence have a larger number of offspring compared to followers on the same patch. Inequality in resources limits the development of hierarchy because it increases the number of offspring leaders produce, and potentially drives all individuals within a group to develop high influence. This effect can be seen in figure 4a, which shows that a higher level of inequality reduces the skewness of the distribution of influence. But this is limited by the competition between leaders. In the presence of multiple leaders, a leader can get a lower share of the resource than followers if the group becomes convinced by another leader during the decision-making process. In this case, the ‘losing’ leaders are further from the final decision because they are harder to convince. However, the fact that hierarchy does not evolve for high levels of inequality shows that this competition is not always enough to stop the increase in number of leaders and the collapse of hierarchy. The second reason explaining the evolution of hierarchy despite inequality is that, even if leaders receive more resources, followers still get a higher amount of resources and offspring than they would in a group without hierarchy, because larger groups produce more resources owing to increasing returns to scale. Figure 4. Average skewness of the distribution of influence α across 5000 generations and across 32 replicates as a function of the level of ecological inequality d and the number of listeners Nl. (Online version in colour.)
We have seen previously that hierarchy reduces the consensus time but it also provides a second main advantage to group organization: it reduces scalar stress. To test the importance of this factor in the evolution of hierarchy, we look at the skewness of the distribution of influence for different values of number of listeners, Nl. Figure 4b shows that high scalar stress, i.e. low number of listeners Nl, leads to the evolution of a more skewed distribution of influence. On the other hand, a low scalar stress, i.e. here represented by a high number of listeners Nl, leads to the disappearance of hierarchy. This result shows that the benefit of reducing scalar stress is a crucial factor in the evolution of hierarchy. This is because scalar stress creates a positive feedback loop by which hierarchy increases its own benefit. On the one hand, an efficient hierarchical organization allows a group to produce a larger amount of resources and hence reach a larger size. On the other hand, hierarchy provides a stronger advantage as group size increases because the cost of organization increases less in hierarchical groups than in acephalous groups. There is a feedback loop between hierarchy leading to larger group size, and larger group size increasing the benefit of hierarchy. Eventually, this feedback loop comes to an end owing to diminishing marginal returns, i.e. other limiting factors than group size. Yet, this feedback loop amplifies the benefit that hierarchy provides to group members and favours its evolution. To summarize, social organization is the equilibrium between two opposing forces, competition within groups where inequality pushes individuals to evolve high influence, and competition between individuals of different groups where efficient group organization pushes most individuals to evolve low influence. Looking closer, hierarchy provides one direct and one indirect benefit [13] to followers compared to individuals in acephalous groups. First, hierarchy provides a direct benefit to followers because it increases the amount of surplus resources produced and thus, it increases the fitness of followers. Second, hierarchy provides an indirect benefit to followers because it increases the group size and hence the amount of resources produced in the following generation. This increases the fitness of followers’ offspring. The contribution of each benefit is hard to distinguish but their role can be examined by investigating the effect of high migration rate, which suppresses population structure and indirect benefits to offspring on the same patch. The electronic supplementary material, figure S6 shows that, considering moderate time constraints, a high migration rate leads to the disappearance of hierarchy at equilibrium. This highlights the importance of the indirect benefit to offspring that remain on the patch in sustaining hierarchy. On the other hand, the electronic supplementary material, figure S7 shows that hierarchy evolves for any migration rate if the the time constraints are high. In this case, the direct benefit is high enough to overcome the cost of inequality in resources. In conclusion, hierarchy can evolve when time constraints are high through the immediate direct benefit of producing extra resources, but the indirect benefit resulting from the feedback loop between hierarchy, group size and scalar stress allows hierarchy to evolve over a much wider range of conditions. Group size and the resultant scalar stress have been proposed as a crucial factor to explain the emergence of hierarchy from previously egalitarian groups [16,17,19]. However, the investigation of this in models of either social dynamics or evolutionary dynamics has been limited so far because a formalization of hierarchy compatible with both types of model was lacking. To fill this gap, we have described group social organization by the distribution of an individual trait, the influence. We have looked at the effect of this distribution on the consensus time using an opinion formation model, and if this distribution can emerge from the evolution of individual behaviours in an evolutionary model. Our results first show that hierarchy reduces the intensity of scalar stress, i.e. the increase of consensus time as group size grows. This benefit emerges solely from the difference of influence between leaders and followers. Second, the results of the evolutionary model show that hierarchy can evolve de novo in the presence of low initial inequality in resources and increasing returns to scale, which are both reasonable assumptions for the small egalitarian groups and societies present during the Neolithic transition [2,36]. This work expands on previous research in social dynamics and evolutionary dynamics by including the role of scalar stress. A previous opinion formation model shows that heterogeneity in individual persuasiveness and stubbornness can affect the time spent to reach consensus [28]. Our findings confirm this result and show that the shortest consensus time is reached for a positively skewed distribution of these traits, as observed in hierarchy. In addition, our results demonstrate that this advantage is greatly correlated with group size. A previous evolutionary model combining opinion formation and evolutionary dynamics showed that hierarchy could evolve when the cost of organization is high, for example as in warfare [37]. However, the hierarchy obtained was unstable and groups often failed to produce any resources. The model presented here demonstrates that scalar stress was a crucial missing factor which, when integrated, leads to the evolution of stable hierarchy and large groups. Our model of informal hierarchy extends previous work on institutionalized hierarchy [19] by showing that this voluntary theory holds even in societies where political institutions are absent, and thus where inequality creates selection pressures towards leader behaviours. Furthermore, the opinion formation model developed here is a first step to move from a benefit of hierarchy that is simply assumed in the model, to a more mechanistic explanation. However, it is still missing some key aspects of group organization e.g. individual knowledge or network structure. Further work should explore how additional factors of group decision-making could amplify or reduce the role of hierarchy in organization. More broadly, our model is in line with theoretical work which proposes that hierarchy emerged because leaders fulfil an important role for the group, e.g. leaders promote cooperation by monitoring the group and punishing the defectors [38]. Importantly, the explanation explored here is not mutually exclusive with previous explanations. Rather, it can complete them. Following the previous example [38], policing in large-scale societies requires efficient decision-making to create the large number of rules [39] and to manage specialized policing forces. Our findings predict that the level of hierarchy, i.e. skewness of the distribution of influence, should increase both with the time constraints on the tasks tackled by the group and with group size. First, there is extensive evidence that human groups tackling tasks with high time constraints such as warfare often switch to a strong hierarchical organization [40]. Second, previous reviews of ethnographic data presents evidence that group size scales with political complexity [16,17,41]. For example, the Inuit population on coastal North Alaska are composed of large groups relying on bowhead whale hunting, a complex coordination task. These populations are thus under high scalar stress and exhibit a strong hierarchy, with leaders who own the hunting equipment deciding the distribution of resources. In comparison, smaller groups of Inuits living on the Mackenzie Delta rely on individual hunting and have a less hierarchical organization [18]. However, the generality of a scenario where the cost of organization, drives the evolution of hierarchy needs to be better estimated with further work exploring the quantitative relationship between individual behaviours, group size and cost of organization, either in laboratory experiments or in real world human groups. Other than scalar stress, our findings predict that low initial inequality in resources and initial increasing returns to scale are necessary for the origin of hierarchy. Much anthropological evidence shows that inequality in resources was strongly limited in pre-hierarchical societies because of the absence of food storage technologies preventing leaders from building up a personal surplus of resources; and the absence of coercive institutions e.g. dedicated armies and tax collection [2]. Increasing returns to scale is commonly observed in modern collective actions and results from synergistic interactions between individuals, such as division of labour and specialization [42]. Archaeological evidence suggests that agriculture could have provided Neolithic society with such scalable means of production [43]. In political sciences, the ‘iron law of oligarchy’ proposes a comprehensive scenario for the emergence of hierarchy and inequality [1]. Our model, combined with previous research [19,44], shows that an evolutionary iron law of oligarchy is a plausible scenario to explain the transition to hierarchy. Expanding human groups switch to hierarchy by evolution of individual behaviours or by group decision [19] to limit the costs of large-scale organization. Later on, leaders use their influence to bias the distribution of collective benefits and costs towards their own interests [44]. Once a few individuals have monopolized economic power and political power, they can then use these advantages in order to sustain their domination [45]. The main benefit of this theory is to provide a common explanation for voluntary and coercive hierarchy. The two opposing sides of hierarchy emerge from the same mechanism, consensus decision-making, which is quicker but biased when the distribution of influence is skewed. Although the iron law was initially proposed to explain human social organization in the post industrial revolution era, our model suggests that its explanatory scope might be wider than first believed. For example, there is evidence that non-human species also use consensus decision-making to coordinate [46]. While it is hard to draw conclusions about other species with the current model, further work could tailor the model to investigate the emergence of leadership in these species. The code is available online at ‘https://github.com/CedricPerret’ in the project ‘ConsensusMod’ for the opinion formation model, and in the project ‘EvolLeadMod’ for the evolutionary model. C.P. and S.T.P. designed the study; C.P. carried out the research; C.P., E.H. and S.T.P. wrote the manuscript. The authors declare no competing financial interests. This project is part of a PhD funded by Edinburgh Napier University. FootnotesElectronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.4973075. References
Page 20Neotropical rainforests are among the most diverse global biomes [1,2], and understanding the origins and drivers behind this exceptional diversity has long been a central challenge for biologists (e.g. [2,3]). Generally, the diversification of Neotropical biota has been explained by past climatic oscillations or intense landscape alterations over time. For instance, the uplift of the Andes mountains and hydrologic changes (e.g. the evolution of the Amazonian drainage system) are historical landscape alterations related to the diversification of Neotropical biota [4,5]. Between 23 and 10 million years ago (Ma), the western Amazonia was covered by a large, long-lived lake named the Pebas system, which grew in size to one million square kilometres in the Middle Miocene (by approx. 16 Ma) [4,6,7]. Notably, the diversification age of several modern Neotropical taxa coincides with these intense landscape changes in the Amazonian region [4]. Moreover, Amazonia is considered as the source of much Neotropical biodiversity [2]. However, empirical studies seeking to understand the processes underlying biological diversification in the Neotropical biota generally are focused on model vertebrates (e.g. birds, frogs, mammals, and squamates) (e.g. [2,3,8,9]) and plants (e.g. [2,10–14]), with limited attention paid to insects and other components of this rich biota. The integration of niche specialization, rapid diversification, extinction, and constant species accumulation is considered to be crucial in explaining diversity in the Neotropics [15]. Moreover, contrasting hypotheses have been suggested in order to explain the extraordinary diversity of the Neotropical region. The first is called the ‘museum model of diversification’, whereby the constant accumulation of diversity has occurred in Amazonia over the past 30 million years (Myr) due to a stable tropical climate [16]. Under this model of diversification, a steady accumulation of species is expected with low extinction rates, constant diversification rates over time, and larger geographical range sizes that correlate with evolutionary persistence. By contrast, the ‘cradle model of diversification’ suggests that geologically recent events triggered large-scale climatic and geographical changes resulting in rapid accumulation of species via high speciation rates [17]. Therefore, under this model of diversification distinct shifts in rates over time is expected with a recent increase in diversification rates and the geographical region is acting as a centre of origin for species diversity. Polistine wasps are a highly diverse insect group in the Neotropics and have proven to be an excellent model for the study of social behaviour evolution because they show different degrees of caste differentiation [18]. These organisms also are useful for biogeographic investigations [19,20] because they are sensitive to climate changes [20,21] and geographical barriers [20,22]. Within Polistinae, swarm-founding social wasps (also called paper wasps) (Hymenoptera, Vespidae, Polistinae, Epiponini) are a monophyletic tribe of approximately 246 described species within 19 genera. These insects are endemic to the Neotropical region, except for three species of Agelaia, two species of Brachygastra, two species of Parachartergus, two species of Polybia, and one species of Synoeca that are found in southern regions of the United States and/or Mexican plateau [23,24]. They exhibit remarkable social characteristics, such as cyclic oligogyny (variable number of functional queens), complex nest architecture, swarm reproduction, subtle morphological difference among castes, and alternative modes of caste determination (i.e. during immature development or after adult emergence) [25]. Previous investigations of the phylogenetic relationships of Epiponini wasps have been conducted with sparse taxon sampling and are derived from studies focusing on higher-level relationships within the Polistinae group. These studies have used morphological data (e.g. [26–28]), a combination of morphology and a few molecular markers (e.g. [29,30]), and most recently approximately 380 genomic loci from 19 epiponine taxa within 18 genera [31], but they have revealed conflicting topologies for some clades of Epiponini. Hence, the sparse sampling within Epiponini and resulting phylogenetic incongruences hamper the robust reconstruction of their evolutionary history. Here, we used massively parallel sequencing of ultraconserved elements (UCEs) to reconstruct a robust phylogenetic tree, infer divergence times, diversification rates, and the biogeographic history of Epiponini wasps. UCEs are a group of abundant nuclear markers distributed throughout the genomes of most organisms [32]; notably, these markers outperform traditional multi-locus approaches [33,34], they can be collected from historical museum specimens for phylogenetics and population genetics [35,36], and they have been successfully employed for phylogenomic studies on a variety of hymenopteran groups [37–41]. Although these genomic elements are highly conserved, hence carrying little phylogenetic information, their flanking regions increase in variable sites as the distance from the core UCE increases, making UCEs excellent markers to study evolutionary relationships across variable timescales [3,42]. Equipped with a phylogenomic approach to investigate the evolutionary history and macroevolutionary dynamics of Epiponini wasps, we tested which diversification model (e.g. ‘cradle’ or ‘museum’) could explain the diversity of the group. We addressed the following questions: (i) was the diversification of Epiponini lineages gradual and ancient (favouring the ‘museum’ model) or more recent and rapid (according to the ‘cradle’ model)? (ii) Is Amazonia the source of Epiponini diversity in the Neotropics with subsequent dispersals across the Neotropics or the group originated outside of Amazonia or even the Neotropics and dispersed to this region before radiating? We sampled 115 Neotropical swarm-founding wasp specimens, representing all 19 genera and 109 of 246 currently described species (see electronic supplementary material, table S1, Appendix S1). We also included 16 other species of Vespidae and one distantly related taxon (Rhopalosomatidae) as outgroups for our phylogenetic analyses. For nine species, we downloaded raw sequencing reads from the Dryad Digital Repository [43]. We extracted DNA from 65 pinned museum specimens ranging in age from 1921 to 2006 and 58 specimens which were field collected and immediately preserved in ethanol (for detailed information see electronic supplementary material, table S1, Appendix S1). The thorax (for ethanol preserved samples) or two legs (for pinned museum specimens) were removed from each sample and total DNA was extracted using a DNeasy Blood and Tissue Kit (Qiagen, Valencia, CA, USA) following the manufacturer's protocol, except the samples were soaked in Proteinase K overnight and the total DNA was eluted in 130 µl ddH2O instead of the supplied buffer. Specifically for pinned museum specimens, before DNA extraction we adopted the recommendations suggested by Blaimer et al. [35] as follows: the tissues were washed in 95% ethanol to remove dust accumulated on the wasp legs and, after evaporation of the ethanol (by drying the tissue on a clean Kimwipe™), the samples were placed in a freezer for at least 6 h before the DNA extraction process. We employed a targeted sequencing approach to collect phylogenomic data from UCE loci [42,44]. For UCE enrichment, we used an RNA bait library for Hymenoptera targeting 2590 loci [40]. The laboratory protocol is outlined in electronic supplementary material, Appendix S2. We performed all bioinformatics steps, including read cleaning, assembly, alignments, and descriptive statistics using the Phyluce v.1.5 software package [45] (see electronic supplementary material, table S1 in Appendix S1 for all sequencing and assembly descriptive statistics). We cleaned and trimmed FASTQ files for adapter contamination and low-quality bases using Illumiprocessor [44] based on the package Trimmomatic [46] and assembled contigs using Trinity v.r2013-02-25 [47]. Subsequently, we used several Phyluce scripts to identify and extract assembled contigs representing enriched UCE loci from each specimen. We used MAFFT v.7.221 [48] and Gblocks v.0.91b [49] to align and trim UCE loci, respectively. We filtered the aligned dataset for taxon occupancy (percentage taxa required to be present in a given locus) (electronic supplementary material, table S2, Appendix S1 for all matrix statistics). Additional detail on matrix preparation is outlined in electronic supplementary material, Appendix S2. We examined the effects of phylogenetic inference methods, minimum number of UCE loci per specimen, and data partitioning in the phylogenetic trees. For phylogenetic inference, we compared maximum likelihood (ML), Bayesian inference (BI), and species tree (ST) approaches (electronic supplementary material, tables S3, Appendix S1). For concatenated ML, we performed analyses with four different partitioning schemes using RAxML v.8.2.11 [50]: unpartitioned, partitioned by locus, partitioned using PartitionFinder v.2.1.1 (PF) [51] with the hcluster algorithm [52], and partitioned using PF with the rcluster algorithm [52]. For each analysis, we executed 100 rapid bootstrap inferences, best tree search (‘-f a’ option), and we used the GTR+Γ model of sequence evolution (for both best tree and bootstrap searches). For BI using only unpartitioned data matrices, we used ExaBayes v.1.5 [53]. For all BI, we executed two runs in parallel with 500 000 generations, each with four coupled chains (one cold and three heated chains). We assessed burn-in, convergence among runs, and run performance by examining log files with the Tracer v.1.6 [54]. We computed consensus tree using the ‘consense’ utility, which is included in the ExaBayes package. We performed ST using the program ASTRAL-III v.5.5.9 [55,56]. First, we used RAxML to generate unpartitioned gene tree for each UCE locus. We ran the ASTRAL-III analysis with 100 multi-locus bootstrap replicates. Additional details on all phylogenetic analyses are outlined in electronic supplementary material, Appendix S2. We used BEAST v.1.8.4 [57] to generate a time-calibrated phylogenetic tree for Neotropical swarm-founding social wasps. We used ‘clocklikeness’ scores to sample 100 UCE loci and a fixed tree (see electronic supplementary material, Appendix S2 for details). We used four fossil calibrations to constrain the dating analysis (electronic supplementary material, table S4, Appendix S1). We performed four independent BEAST runs, each with 100 million generations, and sampled every 10 000 generations. For the clock model, we selected uncorrelated lognormal, for the substitution model we used GTR+Γ, and for the tree prior we used a birth–death model. We assessed burn-in, convergence among runs, and run performance by examining log files with Tracer v.1.6. After removing burn-in, we combined trees using LogCombiner and generated a maximum clade credibility tree using TreeAnnotator (both programs included in BEAST package). Additional details are outlined in electronic supplementary material, Appendix S2. We inferred the ancestral range of the Neotropical swarm-founding wasps applying a dispersal extinction cladogenesis (DEC) model [58] implemented in the R package BioGeoBEARS [59] using a variety of constraints. Because both the standard biogeographic model-selection framework and the model describing founder event speciation (the ‘j-model’) may be biased [60], we chose to apply the DEC model alone since it is robust to complex biogeographic scenarios [61]. As a comparison, we also performed a DEC analysis in RASP v.4.2 [62]. For these analyses, we used the BEAST maximum credibility tree pruned to include only Epiponini wasps. For each terminal species, we coded for the presence/absence in the following areas: (A) Nearctic, (B) Andes and Mesoamerica, (C) northern Amazon, (D) south-western Amazon, and (E) eastern South America. Additional details are outlined in electronic supplementary material, Appendix S2. Additionally, we used the Bayesian program BAMM (Bayesian analysis of macroevolutionary mixtures) v.2.5 [63–66] and the R package BAMMtools [67] to identify diversification rate shifts on the phylogeny. For input into BAMM, we used the BEAST maximum credibility tree with all outgroups pruned. We used BAMMtools to select appropriate priors for the BAMM analysis. We then ran the BAMM analysis with four chains for 100 000 000 generations and sampling event data every 10 000 generations. Finally, we explored the BAMM output using BAMMtools and selected the best rate-shift configuration by assessing posterior probabilities. Additional details are outlined in electronic supplementary material, Appendix S2. Moreover, we investigated the temporal accumulation of Epiponini lineages assessing a lineage-through-time (LTT) plot using the R package phytools 0.6-44 [68] with the maximum credibility tree obtained in the BEAST analysis with all non-Epiponini taxa pruned. The sequencing of UCE loci resulted in an average of 2.36 million reads per sample (electronic supplementary material, table S1, Appendix S1), and an average of 23 845 contigs with a mean length of 419 base pairs (bp) that were assembled by Trinity after adapter- and quality-trimming, and showing an average coverage of 11X. The average of n50 and n50_size were 7735 and 282, respectively. Considering contigs representing UCE loci, we recovered an average of 782 UCE loci. Notably, we successfully recovered a considerable number of UCE loci from several pinned museum specimens older than 30 years (for details see electronic supplementary material, table S1, Appendix S1), despite the degradation of DNA in historical museum specimens. Across levels of taxon completeness, our concatenated supermatrices exhibited a range of 12–2097 UCE loci, a range of length of 2595–549 840 bp, and a range of 605–79 936 informative sites (see electronic supplementary material, table S2, Appendix S1). When we compared our different partitioning schemes (unpartitioned, by locus, and both PF hcluster and rcluster algorithms), the phylogenetic results from partitioned by locus ML showed the best log likelihood and clocklikeness scores (see electronic supplementary material, table S3, Appendix S1), an expected result since more parameters are used with partitioning by locus. Our multiple phylogenetic analyses showed that the interspecific relationships for the subfamily Polistinae could be described as follows: (Ropalidiini + (Mischocyttarini + (Polistini + Epiponini))) (figure 1). Thus, our analyses supported with a high bootstrap score (greater than 95%) the cosmopolitan genus Polistes (Polistini) as sister to Epiponini. The only exception was when we performed ST analyses that exhibited Mischocyttarini as the sister of Polistini, but with weak support. Figure 1. Time-calibrated phylogeny of Neotropical swarm-founding social wasps and closely related groups. The tree topology was recovered by analysing the Epiponini-102T-200 L-F50 matrix using RAxML (partitioned by locus; 950 UCE loci; 257 561 bp). We estimated divergence dates using BEAST with the 100 best (based on clocklikeness score) UCE loci, fixed topology, and four-node calibrations (see electronic supplementary material, table S4, Appendix S1). Scale bars next to wasp photos correspond to 5 mm. (Online version in colour.)
All phylogenetic analyses (ML, BI, and ST) recovered a highly resolved phylogeny for Epiponini with most nodes showing 100% bootstrap scores (ML and ST) and posterior probabilities (BI) (figure 1, electronic supplementary material, table S6 in Appendix S1, and phylogenetic trees in electronic supplementary material, Appendix S3). There was no topological difference between ML and BI analyses and almost all nodes received maximum support, but ST analyses exhibited a few topological differences when compared to ML and BI. Some relationships (considering all methods) recovered here conflict with previous studies using morphological, behavioural, and molecular (Sanger sequencing) data [27–29], but were very similar to a study using genomic information [31]. As found by Piekarski et al. [31], we inferred Angiopolybia as sister to all remaining swarm-founding social wasps, rather than Apoica as previously suggested [26,28,29]. Agelaia and Apoica are sister genera, and Epipona is the sister of Synoeca, rather than Polybia. All phylogenetic analyses recovered a monophyletic Polybia with a deep split into two major clades, but most of the subgenera are paraphyletic. Relationships among all epiponine genera were recovered with high support. We estimated for the first time the timescale of the evolution of Epiponini. The tribe originated during the Eocene around 44.9 Ma (95% of high posterior density (HPD): 37.6–51.3 Ma) (figure 1 and electronic supplementary material, Appendix S3). The crown-group age estimates for several genera occurred during the Miocene, including: Agelaia (15.8 Ma; 95% of HPD: 13.6–19.4 Ma), Angiopolybia (19.9 Ma; 95% of HPD: 8.5–34.1 Ma), Apoica (11.7 Ma; 95% of HPD: 5.5–20.6 Ma), Brachygastra (13.6 Ma; 95% of HPD: 9.1–19.1 Ma), Charterginus (12.1 Ma; 95% of HPD: 6.8–18.2 Ma), Chartergus (9.7 Ma; 95% of HPD: 3.8–17.2 Ma), Chartergellus (5.4 Ma; 95% of HPD: 2.0–10.3 Ma), Leipomeles (8.7 Ma; 95% of HPD: 3.0–16.1 Ma), Parachartergus (13.7 Ma; 95% of HPD: 8.5–19.9 Ma), Protopolybia (15.5 Ma; 95% of HPD: 9.9–21.3 Ma), Pseudopolybia (15.6 Ma; 95% of HPD: 6.3–25.5 Ma), and Synoeca (5.5 Ma; 95% of HPD: 3.0–9.0 Ma). The crown-group age estimates of Epipona (2.3 Ma; 95% of HPD: 0.6–4.4 Ma) and Metapolybia (3.3 Ma; 95% of HPD: 1.5–5.9 Ma) were more recent, during the Pliocene. The diversification of all sampled extant Epiponini species occurred during the Miocene and Plio-Pleistocene (figures 1 and 2). Figure 2. Biogeographic history and diversification dynamic of Neotropical swarm-founding social wasps. (a) Ancestral ranges inferred using the R package BIOGEOBEARS with the DEC model. Coloured squares indicate current or putative ancestral geographical ranges. We used the following ranges: (A) Nearctic, (B) Andes and Central America, (C) northern Amazon, (D) south-western Amazon, and (E) eastern South America. The upper panel is adapted from Hoorn et al. [4]. (b) Net diversification rate through time for Epiponini. Blue shaded areas represent posterior probability distributions of rate estimates. (c) Extinction rate through time for Epiponini. Blue shaded areas represent posterior probability distributions of rate estimates. (d) Phylorate showing diversification rate. Colours of branches indicate the mean evolutionary rate (relative rates from blue (slower) to red (faster)). (e) Lineages-through-time plot for Epiponini. Red dashed line marks a relatively gradual accumulation of lineages in Epiponini. (Online version in colour.)
Our results from DEC biogeographic analyses using BioGeoBEARS and RASP are provided in figure 2a and electronic supplementary material, Appendix S3. Epiponine wasps likely evolved from an Amazonian ancestor, and this area was indicated in our biogeographic reconstructions as having played a major role in the diversification of epiponine wasps, as supported by the fact that all genera have an Amazonian ancestor and its highest taxonomic diversity is also in this geographical region. Additionally, some lineages are inferred to have dispersed from the Amazon to Central America and eastern South America. Moreover, our results show no major shifts in diversification rates for Epiponini over the past 40 Ma, but there was a slight increase in diversification rate over time (figure 2b,c). Extinction rates appear to have remained low over the evolutionary history of Epiponini (figure 2d) and lineage accumulation was gradual (figure 2e). Our extensive sampling of taxa and UCE loci not only produced a strongly supported phylogeny for Epiponini but also provides a detailed window into the evolution of this insect group. Epiponini originated in the Eocene and most diversification events probably occurred gradually within the Amazon region, especially during the Neogene (Miocene and Pliocene) when Andean mountain building caused large-scale climatic, geological, and hydrological changes in South America [4,15] (figures 1 and 2). For example, the intensification of the Andean uplift and subsequent changes in the Amazonian landscape generated the formation of the Pebas system—a large wetland of shallow lakes and swamps in western Amazonia between 23 and 10 Ma [4] (figure 2)—which favoured the diversification of several South American lineages [69]. The gradual accumulation of lineages and no major shifts in diversification rates over the past 30 Myr, as indicated for Epiponini, is a pattern also revealed for the Neotropical plant genus Brownea (Fabaceae) [14] but different from that observed for the Neotropical bellflowers (Campanulaceae: Lobelioideae) and Inga plants (Fabaceae: Mimosoideae) that experienced rapid radiation within the last 5 Myr [11,12]. Our biogeographic reconstructions suggest the Amazonian region as the major source of Epiponini diversity, which can be explained by the highest taxonomic diversity of the tribe in this geographical region and recent colonization events to other geographical areas (e.g. eastern South America, and Central and North America). An Amazonian ancestral distribution during much of the evolution of the tribe is also supported by the fact that all Epiponini genera also have an Amazonian ancestor. Additionally, we can infer recent colonization events to other Neotropical regions in some Epiponini lineages. For instance, the Pebas system may have acted as a barrier preventing Epiponini from reaching Central and North America, but during this same period some lineages colonized eastern South America (figure 2a). Only after the end of the Pebas system during the late Miocene did these wasps colonize Central and North America (figure 2a). Thus, we propose the Amazon Forest as the major centre of origin for Epiponini wasps, as previously suggested for some Neotropical Polistinae groups [70,71] and plant taxa (for example, the Brownea clade) [14], with subsequent migration events across the Neotropics. Phylogenetic hypotheses for the four tribes of Polistinae are historically controversial. Phylogenies based exclusively on morphological and behavioural data suggest that Polistini is the sister taxon to the remaining lineages of Polistinae [26,72]. However, molecular phylogenies [29,31,73] and our phylogenomic analysis using UCE data showed Ropalidiini as the first lineage of Polistinae and recovered with high support Polistini and Epiponini as sister groups. All our analyses recovered a consistent, highly supported phylogeny for Epiponini wasps; the only exception was ST analysis which exhibited slight topological differences (but with low support) when compared to ML and BI. Incongruences between ST and concatenation methods have been explained by the presence of incomplete lineage sorting (ILS) [74] or gene flow in the form of hybridization or introgression [75,76]. Despite demonstrations of hybridization events in other hymenopterans such as bees [77] and ants [78,79], we currently do not possess clear evidence of such events between lineages of Epiponini. Our study now provides the first comprehensive phylogeny of Neotropical swarm-founding social wasps using both genomic data and complete genus sampling. Our results strongly support the monophyly of Epiponini and all genera within the tribe. Regarding the systematics of the tribe, a striking result in our study, and also found by Piekarski et al. [31], is the strong support for Angiopolybia as sister to the other lineages of Epiponini as well as Agelaia and Apoica as sister taxa. These results conflict with previous studies [27–29], but as suggested by Noll et al. [80] the shape difference of queens may be a synapomorphy for Agelaia and Apoica. A second interesting result is strong support for the monophyly of Polybia, with previously proposed subgenera as paraphyletic, suggesting the need for future work on the systematics of Polybia. Carpenter et al. [81] found Polybia as monophyletic with weak support and Pickett & Carpenter [29] did not recover Polybia as a monophyletic group. In summary, our results provide a robust phylogeny for Polistinae and detailed information on the diversification, macroevolutionary dynamics, and historical biogeography for Epiponini. We suggest that the group diversified according to the ‘museum’ and the ‘cradle’ models of diversification as also proposed for ants [82] and the cladogenetic events occurred mainly in Amazonia. These conclusions are supported by several results: (i) the diversification of many genera and all sampled extant species occurred during the Miocene and Plio-Pleistocene, (ii) the Amazon region remained the dominant ancestral distribution during much of the tribe's evolution, and (iii) we inferred no major shifts in diversification rates for Epiponini over the past 40 Ma but with a slight increase towards the present, and lineage accumulation was gradual, probably with a low extinction rate. The spatio-temporal pattern recovered herein for Epiponini reflects an evolutionary history most likely concurrent with climatic and landscape changes during the Miocene and Pliocene and establishes the Amazonian region as the major source of Epiponini diversity. Additionally, we suggest that further studies accounting for diversification rates and variables such as paleotemperature or Andean paleoelevation are needed to determine the impact of past geological and temperature changes in the diversification of social wasps of Amazonia. Data available from the Dryad Digital Repository: https://doi.org/10.5061/dryad.59j3k8p [83]. R.S.T.M. and S.G.B. conceived and designed the study. R.S.T.M. carried out the molecular laboratory work and performed the analyses. R.S.T.M. drafted the manuscript. M.W.L. helped carry out the molecular laboratory work and revised the manuscript. S.G.B. helped with data analysis and manuscript revision. All authors gave final approval for publication. The authors have identified no conflict of interest to disclose. R.S.T.M. is thankful to the São Paulo Research Foundation (FAPESP) by postdoctoral fellowships and grants nos. 2015/02432-0 and 2016/21098-7 and the Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) by grant no. 431249/2018-0. S.G.B. received research support from the U.S. National Science Foundation grant no. DEB-1555905. We are grateful to Matthew L. Buffington and the Department of Entomology at the National Museum of Natural History (https://entomology.si.edu) for the access to the high-quality imaging system; Alexandre Somavilla for contributing specimens to this study; Daercio Lucena, Fabiano Stefanello, and José Amilcar Tavares Filho for assistance in the field; and Eduardo A. B. Almeida for discussions about biogeography and suggestions and comments that improved the manuscript. We also thank James Carpenter (curator of Hymenoptera, American Museum of Natural History) for providing access to the museum collection and suggestions and comments that improved the manuscript. We thank three anonymous reviewers for their valuable suggestions and comments. All laboratory work was conducted in and with the support of the L.A.B. facilities of the National Museum of Natural History, Smithsonian Institution. The computations performed for this paper were conducted on the Smithsonian High Performance Cluster (SI/HPC), Smithsonian Institution. https://doi.org/10.25572/SIHPC. FootnotesElectronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.4971215. References
Page 21In ecology, the term ‘refugium' is used to describe regions that facilitate the temporal and spatial resilience of biological communities over evolutionary timescales or from past climate change [1,2]. Extreme climatic events (ECEs) can generate changes in species distributions, ecosystem structure and functioning [3–7] and are predicted to increase in magnitude and frequency because of climate change [8,9], and thus, taxa are being impacted in ecological timescales. To address the response of taxa to such ecological catastrophe, we have adopted the term ‘refuge' to define spatial or temporal facilitation of environmental conditions or biotic interactions [2], that may enable the persistence of a species and/or communities in ecological timescales (years–decades) [2]. Climate change models have predicted shifts to the distribution of numerous species in terrestrial and marine ecosystems in response to climate change-related pressures [10–12]. However, species distributions are more affected by the local environment which in some cases may provide refuge and allow species to survive the climatic stress [13,14]. In the light of the increase in extreme climatic events (such as atmospheric and marine heatwaves (MHWs), droughts and wildfires) driven by climate change and their catastrophic consequences in marine environments [15–18], identifying refuges has become a research and conservation priority [2], because they have the potential to prevent the extinction of local populations associated with extreme disturbance events. As such, deeper marine habitats have been identified as potential refuges for shallow-reef species [19–22]. Yet, the logistical difficulties of surveying deeper communities hinder our understanding of how these communities respond to extreme climatic events and whether their response varies from shallower reefs. Biogeographic transition zones are biodiversity hotspots owing to the overlap of taxa present at the edge of their biogeographical distribution (tropical and temperate taxa). These zones are particularly vulnerable to the disturbances of extreme climatic events as many species already live at the physiological limits of their distribution [23]. Oceanic MHWs are extreme climatic disturbances that are predicted to increase in frequency and intensity owing to climate change [8]. MHWs are defined as extended periods of anomalously high sea surface temperatures [24] which have already resulted in devastating effects on coastal marine communities characterized by widespread mortalities of invertebrates [3,25], seagrasses [26], coral, (associated with coral bleaching) [27], range contractions of habitat-forming species [28], and changes to community structure and ecosystem function [3,23]. MHWs have been documented across the globe: in the Mediterranean [29]; Australia [30]; northwestern Atlantic [31] and in the northeastern Pacific [32]. Understanding the response of marine ecosystems to MHWs is a key to predict their response to future climate change. Moreover, the ecosystem recovery from these impacts is variable and depends on processes such as population connectivity, fluctuations in fecundity, post-settlement success and altered species interactions. In Western Australia during the summer of 2010/11, an extreme MHW superimposed over a general trend of ocean warming placed a global marine biodiversity hotspot at catastrophic risk [33]. This event was characterized by record high-temperature anomalies that extended across 12° latitude (Ningaloo Reef at 22° S to Cape Leeuwin at 34° S), up to 200 km offshore and down to depths of 50 m [33], with highest anomalies of +5°C around the central coast. The response of benthic marine communities to this extreme event is well documented for shallow habitats (less than 15 m depth). Kelp beds were lost across approximately 2300 km2 causing a 100 km range contraction. Kelps and other macroalgae were replaced by less complex turf-forming algae and the recovery of kelp suppressed owing to the grazing pressure driven by an increase in tropical herbivorous fishes [23]. A staggering 1069 km2 of the 4366 km2 of seagrass meadows in Shark Bay were lost during and immediately after the 2011 MHW, resulting in significant ecosystem-wide changes [26,34], although recovery of 125 km2 of meadows has occurred since 2014 [34]. At Ningaloo Reef, bleaching was observed in 79–92% of the coral cover [35] and at the Houtman Abrolhos Islands, the bleached coral was reported to be 6–42% varying across sites [36]. Despite all the evidence on the effect of MHWs on shallow marine ecosystems (less than 15 m), the response of deeper (greater than 15 m) benthic communities is often not documented and thus poorly understood. As marine ecosystems continue to be degraded, identifying regions that can be a refuge for key species has become a priority for management and conservation [37]. Deeper marine reefs were first identified as refuges in the context of tropical coral reef ecosystems as research on mesophotic coral reefs increased [20,21,38]. Species living in deeper habitats may benefit from a higher chance of survival from extreme environmental events owing to the buffering effect of depth [39,40]. Nonetheless, refuge habitats also need to share similar species with the habitat they are providing refuge for (in this case, shallow reefs). As a result, in the case of benthic species such as coral and algae, deep refuges are thought to be constrained to the upper regions of the mesophotic zone, typically shallower than 60 m [41]. However, most research on the ecology of deeper habitats and their potential role as refuges for shallow-water species has been focused in tropical coral reef ecosystems [41,42], while subtropical and temperate regions remain understudied. In order to address the potential for deep habitats to act as refuges for temperate marine benthic communities, we aim to recognize if depth moderates the response of benthic communities before, during, and after an extreme MHW. We firstly characterize benthic community composition along a temperate to tropical biogeographic transition zone and across a depth gradient (15 m–40 m). Second, we describe how the community composition of deep benthic habitats was affected by the MHW (2010–2011) and how the response of deeper reefs compared to shallower reefs. Finally, we identify the key macroalgae species across the biogeographical transition zone and evaluate how their abundance changed with increasing depth and after the 2011 extreme MHW. We argue that deep benthic habitats in the biogeographic transition zone of Western Australia acted as refuges from the 2011 extreme MHW, as suggested by the reduced response of community composition and foundation species to this event. If depth helps foundation species moderate their response to future ECEs, these habitats may constitute refuges from future MHWs and from climatic change. Benthic images were obtained from surveys conducted at permanent monitoring sites that were established by the Australian Integrated Marine Observing System (IMOS) initiative at the Houtman Abrolhos Islands, Rottnest Island and Jurien Bay [43] (figure 1). Surveys were conducted with an autonomous underwater vehicle (AUV) which records down facing georeferenced stereo image pairs, along with a suite of physical parameters including multibeam bathymetry, temperature, salinity, and chlorophyll a. Surveys were completed at two sites in each location, with an additional site at Abrolhos to account for a missing 25 m site, and at three depths: 15, 25 and 40 m (figure 1). Within each site, three replicate ‘grids' were surveyed at the beginning of the monitoring programme in 2010. Each replicate ‘grid' comprises a 625 m2 area (25 × 25 m) of the seafloor surveyed by conducting parallel overlapping 25 m long transects across the seafloor. Grids at each location were located 50–200 m apart for spatial independence. Subsequently, surveys were repeated at every location every year until 2013, and the site Snapper Bank was added to Abrolhos Island. Only Abrolhos Islands were surveyed in 2014 and surveys were repeated at Abrolhos Islands and Rottnest Island in 2017 (electronic supplementary material, table S1). Repeated surveys aimed to assess the same three grids initially established, though on some occasions only one or two grids were surveyed owing to unfavourable weather conditions or equipment malfunction. More than 1000 stereo image pairs were captured by the AUV at each grid, but only 30 non-overlapping images (approx. 4 m2) per grid were randomly subsampled and processed as power analyses have indicated minimal improvement in detectable sizes for image replication above 30 [44]. Across all locations and times, this gave a total of 5970 images for analysis. Figure 1. Locations of benthic surveys in Western Australia (Houtman Abrolhos Islands, Jurien Bay and Rottnest Island). The replicate grids are shown with respective depths, red: 15 m, yellow: 25 m and blue: 40 m. (Online version in colour.)
The three locations surveyed in this study comprise the transition zone between the subtropical and temperate coast of Western Australia. The locations of Houtman Abrolhos Islands (28°43′S), Jurien Bay (30°29′S) and Rottnest Island (32°00′S) were chosen for long-term monitoring as they have been identified as key indicator regions because of their ecological importance as biogeographic transition zones and owing to their socio-economic importance in the region [43]. The Houtman Abrolhos Islands to the north form an archipelago 80 km from the mainland. The Abrolhos Islands are unique in their benthic assemblages, where patches of the kelp species Ecklonia radiata can be found coexisting with reef-building hard corals. Jurien Bay is situated between Abrolhos Island and Rottnest Island in a region characterized by inshore lagoons protected by offshore reefs and islands. The limestone reefs are dominated by E. radiata, other macroalgae, some corals and sponges. Rottnest Island in the south is located 19 km from the mainland near the city of Perth. The island is surrounded by complex limestone reefs that are dominated by kelp species E. radiata and include numerous macroalgae species, seagrass meadows and coral. Each image was annotated by classifying the substrate, flora or fauna beneath 20 randomly and digitally overlaid points using Coral Point Count with Excel Extensions (CPCe) [45]. Each point was classified into functional/morphological groups with 104 categories in total, consistent with the Collaborative and Automated Tools for Analysis of Marine Imagery (CATAMI) classification scheme [46]. CATAMI provides a standardized vocabulary for image classification, enabling the compilation of regional, national and global datasets. Species of ecological importance were classified to species level and included E. radiata, Scytothalia dorycarpa and Sargassum sp. It is important to note that this method is poor at quantifying rare taxa or taxa smaller than 5 cm [47]. For each grid, subsampled images were pooled and the analysis of community composition was conducted using the grids as spatially independent replicated units. The multivariate community composition of the region was evaluated with a principal coordinate analysis (PCO) performed with Bray–Curtis similarity matrices based on square-root transformations of the data with a dummy variable (value of 1) used to optimize the year-to-year separation owing to a large number of zeroes in the data. The centroids represent means for each site per location, depth and year resulting from two or three grids. Community composition across depths was also examined at each location by constructing PCO plots. The centroids represent the averages for each depth per year derived from four to six grids. From this, we determined a trajectory of change in community composition in response to the 2011 MHW. In all PCOs, vectors over 0.7 correlation are illustrated to identify the benthic classes that characterize the assemblages. Plots of change in per cent cover following the MHW (2010–2011) and compared to the latest survey (2010-last survey) were generated for E. radiata, turf, encrusting red algae and S. dorycarpa, so we could compare it to reported changes in inshore reefs [3] and to assess whether any recovery occurred. These plots were made by calculating the mean per cent cover of each species (or benthic class, like turf and encrusting red algae) at each grid, per location, depth and year, and then calculating the absolute change in per cent cover from 2010 to 2011 and from 2010 to the last survey (which varied with location, see the electronic supplementary material, table S1). Differences in per cent cover for E. radiata, turf, encrusting red algae and S. dorycarpa for each location and depth, were analysed by one-way analyses of variance (ANOVA) between 2010, 2011 and the year of the last survey, followed by a Tukey-test if differences were significant. When assumptions of normality and homogeneity of variance were violated, a Kruskal–Wallis test was used and Dunn's post-hoc test (electronic supplementary material, table S2). Plots of mean per cent cover of principal benthic categories for each location, depth and year are presented to visualize their change in abundance through time and in response to the MHW. Certain benthic categories are only described for one or two locations, such as coral at Abrolhos Islands. At each location, differences in per cent cover of ecologically important benthic categories were tested between depths and across years with univariate PERMANOVAs, with depth (three levels) and year (six levels) as fixed factors. Data for each location consisted of the per cent cover of individual images, rather than the pooled grid averages used for the multivariate analyses. The tests used 9999 permutations of square-root transformed data and an Euclidean distance resemblance matrix. Three distinct community groups along the subtropical–temperate biogeographical transition zone of Western Australia were evident from the PCO: one for the shallow Abrolhos sites (15 m), one for Jurien and the deeper Abrolhos sites (25 and 40 m), and one for Rottnest Island (figure 2a,b). Figure 2. Principal coordinate analysis (PCO) of variation in benthic community structure at Abrolhos Islands, Jurien Bay and Rottnest Island based on a Bray–Curtis similarity matrix. The first two axes explain 64.1% of the variability in multivariate space. (a) Centroids represent average community composition at all locations for each year at each depth. Black rings indicate the centroids with 60% similarity. (b) Vectors indicating benthic categories with high correlations (Spearman correlation > 0.7) with axes. (Online version in colour.)
The Abrolhos Islands were characterized by a mixed assemblage that varied markedly with depth (figure 2a,b). Shallow sites were dominated by reef-building corals of staghorn, tabulate, massive and foliose morphology and were colonized by the brown algae taxon Sargassum sp., while the deeper sites were either characterized by sparse E. radiata, S. dorycarpa or sand (figure 2a,b; electronic supplementary material, figure S1). Jurien Bay presented a community composition similar to the deeper sites of Abrolhos Islands (25 and 40 m) (figure 2a), characterized by a higher percentage of sandy substrate and fine branching red algae (figure 2b; electronic supplementary material, figures S1 and S4). Rottnest Island's community composition was dominated by large brown macroalgae at all depths, in particular, E. radiata and S. dorycarpa at the shallow sites, with encrusting red algae cover increasing with depth (figure 2a,b). At Abrolhos Islands, there was a trend of greater change across years in community composition at 15 and 25 m, and less at 40 m (figure 3a). The only convergent community composition among years was shown between 2010 and 2017 at the 15 m site at Abrolhos. Following the MHW (2010 to 2011), the 15 m sites of Abrolhos Islands changed in community composition with an increase in bleached coral (approx. 4%) (electronic supplementary material, figure S2) and turf matrix (approx. 20%) (figure 4b; electronic supplementary material, figure S3), and a decrease in fine branching red algae (approx. 11%) (electronic supplementary material, figure S4) and foliose coral (approx. 5%) (electronic supplementary material, figure S5). Minimal change in encrusting red algae cover (approx. 2%) was seen in shallow sites (figure 4c; electronic supplementary material, figure S6) and a decrease in E. radiata (approx. 3%) was observed (figure 4a; electronic supplementary material, figure S7). By 2017, the shallow (15 m) benthic community of Abrolhos Islands appeared to have returned to a state similar to pre-heatwave composition (figure 3a). The 25 m sites showed a response to the heatwave with an increase in Sargassum sp. (approx. 3%), E. radiata (approx. 10%) and seagrass (approx. 5%) (electronic supplementary material, figures S7–S9 respectively). By 2017, community composition had not returned to pre-heatwave conditions, with reduced turf cover (approx. 15%), increased encrusting red algae (approx. 3%), Sargassum sp. (approx. 3%), and seagrass (approx. 3%) (figure 3a). On the other hand, there was minimal benthic community change at the Abrolhos 40 m sites (figure 3a) after the 2011 MHW and between all the years. Figure 3. Principal coordinate analysis (PCO) of variation in benthic community structure at each location based on a Bray–Curtis similarity matrix. Centroids represent average community composition at each location ((a) Abrolhos Islands, (c) Jurien Bay, (e) Rottnest Island) for each year and depth and arrows indicate the trajectory. Thicker arrows show the change in average community composition from 2010 to 2011. Vectors ((b) Abrolhos Islands, (d) Jurien Bay, (f) Rottnest Island) indicate the benthic categories with high correlations with axes (Spearman correlation > 0.7). For Abrolhos, the first two axes explain 76.3% of the variability in multivariate space (a,b). For Jurien Bay, the first two axes explain 66.2% of the variability in multivariate space (c,d). For Rottnest Island, the first two axes explain 67% of the variability in multivariate space (e,f). (Online version in colour.)
Figure 4. Absolute change in mean per cent cover (± s.e.) from 2010 to 2011 (heatwave) and from 2010 to latest survey for E. radiata (a), turf matrix (b), encrusting red algae (c) and S. dorycarpa (d) at each location (Houtman Abrolhos Islands, Jurien Bay, and Rottnest Island) and depth (15, 25 and 40 m). Colours describe an increase (blue) or decrease (red) in per cent cover in comparison to 2010. Significant changes in per cent cover are marked with a star. The grey box indicates the level of change reported for inshore reefs in response to the MHW [3]. The estimates of per cent cover are means of 2–6 grids (approx. 30 images per grid) within each location and depth per year. (Online version in colour.)
The Jurien Bay assemblage also changed across years, with the largest change occurring at the shallower sites (figure 3c) after the MHW. In contrast with the other locations, shallow sites at Jurien bay were more characterized by sand patches and seagrass. Seagrasses at this location showed large declines after the MHW (from approx. 8% to almost 0% cover) and no signs of recovery at 15 and 25 m sites (electronic supplementary material, figure S9). The 25 and 40 m sites at Jurien also showed change after the MHW, mostly characterized by increases in encrusting algae (approx. 5% increase at both depths) (figure 4c). Communities at 25 m seemed to be affected by a separate event to the 2011 heatwave, since community composition was recovered to pre-heatwave conditions in 2012, and by 2013, it changed towards a more turf driven community (with an increase of approx. 20%). The community composition at Rottnest Island also responded to the MHW at the shallow sites (15 m), with a reduced response in deeper sites (figure 3e). At the 15 m sites, there was a decrease in S. dorycarpa and E. radiata (approx. 5%, from 2010 to 2011 for both species) (figure 4a,d) and an increase in encrusting red algae (approx. 5%) (figure 4c). In contrast with the other two locations, community composition at shallow sites in Rottnest Island had not recovered to pre-heatwave conditions by 2017 (figure 3e). Cover of S. dorycarpa continued to decrease (approx. 10% decrease by 2013) and had not recovered by 2017 (figure 4d, electronic supplementary material, figure S10). Moreover, the analysis also identified changes in community composition in deeper habitats (25 and 40 m) that appeared to be a response to a process separate from the 2011 MHW, as they were observed from 2013 to 2017 (figure 3e). These changes were driven by an increase of approximately 5% encrusting red algae in the 25 m sites and an increase of approximately 4% in the cover of sponges and S. dorycarpa at 40 m sites (figure 3e,f; electronic supplementary material, figures S7, S10 and S11). Despite changes in per cent cover of macroalgae following the MHW, these were not at the scale of the changes reported for inshore reefs (figure 4) [3]. Decreases of approximately 30% were reported in the cover of E. radiata owing to the MHW at shallow inshore reefs (figure 4a). We found the largest decrease in E. radiata cover at the deep sites of Rottnest Island to be of approximately 18% following the 2011 MHW and at no location or depth were these changes found to be significant (figure 4a). Turf cover increased by around 10% at shallow, inshore reefs after the 2011 heatwave [3], but we only found a comparable increase at the shallow sites of Abrolhos with a significant increase in turf cover of approximately 15% and at the 25 m sites of Jurien with a significant increase of approximately 8% (figure 4b). Other sites and depths did not show a large increase in turf cover after the 2011 MHW and a large significant decrease in turf cover (approx. 20%) was found in the deep sites of Jurien from 2010 to the last survey in 2017 (figure 4b). The largest decrease in encrusting red algae we observed at the 25 m sites at Abrolhos Islands, with a reduction of approximately 5%, but this was not significant, while in shallow inshore sites the reductions were of approximately 15% (figure 4c). Significant increases in encrusting red algae cover were observed in the 15 m and 40 m sites of Abrolhos (approx. 5% at both depths), and at the 40 m sites of Jurien with an increase of approximately 5% after the heatwave and a total of 10% by the time of the last survey in 2013 (figure 4c). Scytothalia dorycarpa at Rottnest Island showed the largest reduction at shallow sites (approx. 5%), and in the last survey, it presented a significant decrease (approx. 10%) compared to pre-heatwave levels, yet these reductions are small compared to the approximately 40% per cent cover decrease at inshore reefs (figure 4d). At all locations, the variation in per cent cover of each benthic category was significantly different by year, depth and their interaction as indicated in multi-factorial univariate PERMANOVA tests (electronic supplementary material, table S3). Exceptions were sponges at Jurien Bay, which did not exhibit an effect of year or its interaction with depth, and sand at Rottnest Island which showed no effect with the interaction of year and depth. MHWs are expected to increase in magnitude and frequency under climate change predictions, posing a threat to the persistence of numerous marine species [48]. Here, we showed evidence that there are potential refuges in deeper offshore reefs for shallow near shore foundation species in temperate regions, such as E. radiata and S. dorycarpa. These depth refuges add a dimension that has not been considered by many studies of widespread mortality on near shore shallow reefs [15,23]. The catastrophic loss of canopy-forming macroalgae, E. radiata and S. dorycarpa documented in shallow waters (less than 15 m) [15,23] was not shown from deeper offshore reefs between 25 and 40 m off Western Australia (figure 4a,d), supporting our hypothesis that deep water habitats exhibit a buffering effect from extreme climatic events, that allows the persistence of kelp-dominated communities. Deep reefs in the mesophotic zone have been proposed to offer refuge from environmental disturbances [49,50], as suggested for tropical mesophotic coral ecosystems [21,38,51]. Here, we analysed benthic community composition across a subtropical–temperate biogeographic transition zone and found a reduced response to an extreme MHW in deeper reefs (25–40 m), despite a natural shift in community composition from mixed assemblages of tropical corals and kelps in the north (Abrolhos Islands) to a typical temperate community dominated by kelps in the south (Rottnest Island) [52,53]. Moreover, key habitat-forming taxa like E. radiata, and fine branching red algae, were found along the transition zone across all depths. Scytothalia dorycarpa was only found in Rottnest Island but showed small decreases in deep sites compared to in shallow ones. Although we observed decreases in E. radiata and S. dorycarpa which had not returned to pre-heatwave status by 2017, these reductions were minimal compared to the decreases reported for inshore communities following the 2011 MHW [3]. Foundation species persisting in deep reefs could provide a source of propagules for their shallow counterparts to facilitate the recovery of shallow disturbed populations [20], provided they are reproductive and have oceanographic connectivity. Because MHWs are predicted to become more frequent and intense in the future [8,9], deep reef communities may be a key driver of shallow-reef resilience, inasmuch as the frequency and magnitude of future MHWs allows for the recovery of shallow communities. In this study, we found that benthic species in shallow offshore sites (15 m) were less affected by the MHW than what was reported in other studies [3,15,23]. For example, populations of E. radiata and S. dorycarpa suffered catastrophic losses in shallow, near shore reef ecosystems and resulted in a range contraction of approximately 100 km at the warmest edge of their distribution [23,28]. The loss of these key habitat-forming species further resulted in ecosystem reconfiguration driven by an increase in less structurally complex turf-forming seaweeds [23]. However, we did not see this regime shift in E. radiata-dominated communities at any depth or location in this study. These results have implications for the spatial scope of benthic surveys and post-disturbance population or community assessments which take into consideration only the shallowest areas of the reef communities and consequently are focused only on the most susceptible area of the species distribution [3,23,28]. Additionally, the models of seaweed distribution along temperate Australia have shown that under ocean warming predictions there will be a significant poleward shift in distribution, with E. radiata being restricted to the south coast [54]. However, we have shown that habitat-building species living in deeper reefs have the potential to persist, and consequently, the range contractions suggested from modelling may have overestimated the total impact of climate change disturbances by not considering the differential response of deeper communities. While temperature anomalies associated with the 2010–2011 MHW have been identified down to approximately 50–60 m of depth [33,55], we did not detect signs of catastrophic alteration in community composition as documented in shallower habitats (less than 15 m), as far south as Rottnest Island. We gathered sporadic, in situ temperature recordings near our study sites over a 20 year period, which also showed temperature anomalies at 40 m depths during the 2011 MHW (electronic supplementary material, figure S14); however, these data lacked enough replication over time to be used for further analyses. Benthic populations living in deeper reefs may be acclimated to frequent thermal variation owing to the effect of the Leeuwin Current which transports warm water from the tropics along the continental shelf of Western Australia [56] and consequently may have greater influence in deep offshore habitats than in shallow and inshore ones [57]. This high variability in water temperature may give them the capacity to withstand MHWs [58–60]. An additional coping mechanism for deep macroalgae may be enhanced photosynthetic efficiency owing to acclimation to lower light conditions, as opposed to their shallow counterparts which were exposed to high temperatures and higher light. This interaction between light and temperature has been shown in studies of the kelp Laminaria saccharina where adult sporophytes acclimated to high temperature and/or low light required less light to achieve positive net photosynthesis than sporophytes acclimated to low temperature or high light [61]. Furthermore, deeper communities, which are often found offshore, may be uncoupled from other co-occurring stressors that affect shallower, coastal ones. The interaction of multiple stressors in shallow coastal ecosystems has been shown to elicit extreme ecological responses via catastrophic loss of species [62] because the effect is synergistic, where the combined response is greater than the sum of individual stressors [63]. Our study did not find evidence for deep refuges in corals; however, this was because the deeper sites surveyed at Abrolhos Islands did not have substantial (greater than 1%) coral cover (electronic supplementary material, figure S2), so no buffering effect of depth could be inferred. The temperate–subtropical biogeographic transition zone of Western Australia provides a model for understanding the effects of climate change on species distribution driven by an increase in sea temperature. Over geological time scales, the Leeuwin current has undergone periods when it was strengthened and weakened and, consequently, contributed to a highly biodiverse region with species adapted to historic temperature ranges [64]. Climate change projections suggest that this region is a warming hotspot where the rate of warming is in the top 10% globally [9], and the most extreme MHW on record was observed in this region with the 2011 Ningaloo Niño [3,65]. Despite some levels of adaptation to temperature shifts, the consequences of changes in benthic communities are expected to be profound [23]. Nevertheless, in this study, we found that deeper habitats were less affected by the discrete warming event in the 2010/11 MHW, where greater depths depict a more stable community with lower species turnover rates [66]. Yet, our understanding of the processes driving the community dynamics of deeper reefs is still in its infancy, as indicated by the large change we observed in community composition at 40 m sites of Rottnest Island in 2017, mainly driven by an increase in sponges which we were unable to associate with an environmental change or disturbance. Despite the persistence of deep populations in deep habitats after the 2011 MHW, a single event is not enough to confirm the existence of deep refuges. The response of deep communities to future extreme events needs to be evaluated to confirm their role as a resilience mechanism for depth generalist species living in shallow reefs. This highlights the importance of continuous monitoring of benthic habitats at different depths. Furthermore, other ecological processes need to be evaluated across depth to confirm the existence of refuge at depth, such as changes to fecundity, transport of propagules from deep to shallow sites and post-settlement survival at shallower disturbed sites. Moreover, these processes may vary across species and the full definition of deep refuges may only apply to some species. For example, S. dorycarpa at 15 m depth at Rottnest Island did not show signs of recovery despite its persistence at deep sites. This may be related to its susceptibility to warm temperatures, which has shown to decrease settlement densities and post-settlement survival of germlings [67], or possibly to reduced fecundity with depth or unsuccessful transport of germlings from deep to shallow sites, all of these processes are currently unknown. Macrophyte communities are quintessential to temperate reefs, providing valuable ecosystem services worth millions of dollars per year [68]. In Western Australia, the western rock lobster (which is endemic to this region) fishery alone is worth over AU$300 M yr−1 with numerous studies identifying E. radiata as critical habitat for adult lobsters [69]. Further research into deeper communities is required to fully understand their potential to act as refuges for shallow benthic foundation species and the ecosystem services they provide. We suggest that deep benthic marine habitats in temperate Western Australia may play a role in buffering the impacts of a recent extreme MHW on the benthic communities found on the continental shelf and, therefore, have the potential to act as a refuge against future extreme climatic events potentially assisting the recovery of shallow-reef communities. If deep habitats are less affected by future extreme events, in the long term, they could act as refuges from climate change, and the range shifts in offshore reefs may be less extreme than projected for inshore systems. It is also essential that these offshore habitats are studied to assist resource managers, particularly in the planning of marine reserves and future proofing of fishery sustainability. Additional results supporting this article have been uploaded as the online electronic supplementary material. Datasets used in this study are publicly available online at https://figshare.com/articles/Benthic_codefile_Depth_moderates_loss_of_marine_benthic_communities_csv/7789538, for benthic classification code and data, and at https://figshare.com/articles/MARVL3Summer_temperature_averages_and_anomalies-Abrolhos_Jurien_Rottnest/7789544, for temperature data. G.A.K., R.K.H. and A.G.O. conceptualized the study and conducted fieldwork. G.A.K. and R.K.H. did an initial analysis on a subset of the data. A.G.O. conducted the complete data analysis. All authors contributed to interpreting the results and writing of the manuscript. The authors declare that they have no competing interests. G.A.K. and R.K.H. organized funding for AUV surveys (Marine Biodiversity NERP, Australian Research Council Grants (grant nos. DP150104251 and LE13010020) and the Integrated Marine Observing System (IMOS) through the Department of Innovation, Industry, Science and Research (DIISR), National Collaborative Research Infrastructure scheme). Fisheries Research and Development Corporation (FRDC) project no. 2008/013 helped cover costs for 2 of the years. We acknowledge the WA Department of Primary Industries and Regional Development Habitat. FootnotesElectronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.4989884. © 2020 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original author and source are credited. References
Page 22The major environmental consequence of urbanization is the transformation of the temporal (day-night) environment, with largely unexplained health consequences [1]. The evidence, largely accumulated from nocturnal rodents, suggests that artificial light at night (LAN) impacts a wide range of biological functions including sleep [1–3]. Similar LAN-induced negative effects have been found on daily cycles of the activity-rest and melatonin secretion, and on metabolism, reproduction, depression and cognitive performance in several songbirds including Indian weaver birds, Ploceus phillipinus [4]; blackbirds, Turdus merula [5]; great tits, Parus major [6–8]; Indian house crows, Corvus splendens [9]; zebra finches, Taeniopygia guttata [10,11] and tree sparrows, Passer montanus [12]. In particular, female great tits were awake for a greater portion of the night inside the nest-box when it was dimly illuminated [6]. Similarly, free-living blue tits (Cyanistes caeruleus) adjusted their awakening time to the prevailing light condition [13], and Indian house crows showed increased sleep and decreased awakening latency under a LAN environment [9]. In diurnal vertebrates, a crucial adaptation is the daily sleep-wake cycle, with sleep as the night (hence, awake as day) component of the 24 h temporal environment. The daily sleep-wake pattern is governed by the internal circadian clock, which operates in a closed transcriptional-translational feedback loop (TTFL) formed by a set of core clock genes [14]. TTFL is functionally arranged in positive (brain muscle arnt like (bmal1); circadian locomotor output cycles kaput (clock)) and negative (period (per); cryptochrome (cry)) limbs [14]. Furthermore, the orphan nuclear receptor genes (retinoid-related orphan receptors (rors) and reverse transcript of erythroblastosis (rev-erbs)) stabilize TTFL by acting as activator and repressor, respectively, of bmal1 transcription [14]. The disruption of the daily sleep-wake pattern results in sleep deprivation (=sleep debt), which can be assessed by the measurement of the circulating oxalate (oxalic acid) levels [15]. However, there are conflicting reports on the reliability of plasma oxalate levels as a biomarker of sleep debt in birds. Under a LAN environment, plasma oxalate levels were significantly dropped in sleep-deprived adult great tits [8], like in rats and humans [15]; however, the oxalate levels were found elevated in developing great tits [16]. Under a LAN environment, even in the absence of an immune challenge, changes in the brain cytokines parallel the sleep disruption, suggesting their involvement in the regulation of sleep [1]. A recent study found tissue-specific 24 h rhythms under 12 L : 12 D (LD) and its disruption under dLAN in genes coding for the pro- (IL-1β, IL-6) and anti-inflammatory (IL-10) cytokines in the hypothalamus and liver of zebra finches [17]. In particular, however, there is evidence for interleukin-1β (IL-1β) and tumour necrosis factor-α (TNF-α), which are synthesized and released via activated toll-like receptor 4 (TLR4), playing roles as the promoter of sleep and regulator of sleep duration [18]. The involvement (at least partially) of TLR4 in sleep loss is also suggested by the lack of cerebral reaction in TLR4-deficient mice [18]. Reciprocally induced IL-1β and TNF-α activate the nuclear factor-κB (NF-ΚB), which via the nitric oxide pathway (shown by nitric oxide synthase (NOS) activity) promotes sleep by acting directly on the hypothalamic preoptic neurons [19,20]. This involves the CamK2 hyperpolarization pathway: camk2 knockout (KO) shortened, and sik3 knockin (encoded SIK3 protein kinase gene acting downstream to CamK2) enhanced sleep duration in mice [21,22]. Similarly, KO mice for NR3A, a Ca2+-dependent N-methyl-D-aspartate receptor subunit, showed a significantly shortened sleep and disturbed sleep homeostasis [21,23,24]. Likewise, ACHM3 (a cholinergic muscarinic receptor subunit and expressed in the ventrolateral preoptic area) was shown to contribute significantly to sleep-wake homeostasis in pigeons, mainly by the activation of hypothalamic mechanisms that maintain the awake state [25]. The mechanism(s) underlying LAN-induced effects on sleep-wake homeostasis is (are) poorly understood in diurnal vertebrates. To this end, songbirds can be a good experimental system because they share many sleep traits with mammals [26]. Like in mammals, sleep allows birds to recover from daily stress, and to consolidate memory, conserve energy and maintain the body temperature and homeostasis [26–28]. Here, therefore, we investigated LAN-effects on sleep and associated molecular correlates in diurnal zebra finches, which show profound LAN-induced negative behavioural and physiological effects [10,11]. We first performed behavioural assays to monitor the sleep and activity-rest pattern, and assayed plasma oxalate levels as a biomarker of sleep effects in adult female zebra finches that were exposed daily to 12 h light (150 lux) coupled with 12 h of absolute darkness or of dim light (5 lux). Then, we measured the hypothalamic expression of genes involved in the circadian timing (per2, bmal1, reverb-β, cry1, ror-α and clock), and in the promotion of sleep (cytokines: tlr4, tnf-α, il-1β and nos; Ca2+ dependent hyperpolarization pathway: camk2, sik3 and nr3a) and maintenance of the awake state (muscarinic cholinergic receptor, achm3) in diurnal zebra finches. This study was approved by the Institutional Animal Ethics Committee (IAEC) of the Department of Zoology, University of Delhi, India (DU/ZOOL/IAEC-R/2018/03). The experiment was carried out on adult female zebra finches (T. guttata) that were born and raised in our indoor facility under 12 h daily photoperiod (12 L : 12 D; L = ∼150 lux; D = 0 lux; temperature = 24 ± 2°C). For this study, we chose females to avoid any potential sex-dependent response to an altered light environment. We also expected a larger response in female birds, based on a previous songbird study in which, when compared with males, female great tits spent a greater proportion of the night awake in response to the artificial LAN environment [6]. The experimental protocol has been described in detail in a previous publication [11]. Briefly, 72 females (age: 8–10 months) were individually housed (cage size = 42 × 30 × 54 cm) kept in separate light-proof wooden boxes (size = 58 × 52 × 68 cm); hence, birds were isolated such that they could not see and hear their neighbours. After a week of acclimation to 12 L : 12 D, as before, birds were randomly separated in two equal groups. For the next three weeks, half of the birds remained on 12 L : 12 D (LD group: L = ∼150 lux; D = 0 lux); however, for the other half of birds, the absolute darkness of 12 h night was replaced by 5 lux dim light (dLAN group: L = 150 ± 5 lux; D = 5 lux ± 1 lux). We chose 5 lux for dLAN based on the average night-time illumination intensity of three different locations in a 6 km2 area around Delhi University that we had measured, and this was consistent with previous dLAN studies in songbirds [7,9]. The white light illumination was provided by Philips (India, 220 V-240 V) compact fluorescent bulbs emitting a radiance of 220 lumens. Each cage was fitted with two such bulbs, one of which was used for providing dLAN at 5 lux light intensity to birds under dLAN. Reduction in the light intensity at night was achieved by covering the bulb with a black paper sheet having multiple slits. We checked and verified intermittently the light intensity both during the day and night by using the Macam Q203 radiometer. To all birds, food (Setaria italica seeds) and water were available ad libitum. Two perches placed in each cage at unequal heights facilitated the perching activity of the bird within the cage environment. The cage was mounted with a passive infrared motion sensor (DSC, LC100 PI Digital PIR motion detector, Canada) that continuously monitored general activity (movement) of the bird in its cage. The movements collected in 5 min bins for each individual were stored separately in a computerized recording system. The collection, collation, graphics and analysis of activity were done by using the ‘The Chronobiology Kit' software program from Stanford Software Systems, Stanford (USA), as described in several previous publications from our laboratory [9]. Double plotted activity records were graphed and presented as an actogram for each bird, wherein successive days were plotted sideways and underneath to show a better visual illustration of the timing and pattern of 24 h activity-rest behaviour. For statistical analysis of the effect of the light condition of activity behaviour, we extracted activity in 20 min bins for seven consecutive days for each bird. We first averaged activity for 20 min bins over 24 h for each individual, and then calculated and presented the mean (±s.e.) 24 h activity profile. We monitored the birds' postures in the cage by using a night vision camera and recorded it as sleep or awake state (figure 1), as per a behavioural assay standardized and used in other bird studies [6,9,13,29]. Videographs were analysed for sleep and awake state by using Observer XT 10 software from Noldus Information Technology (Wageningen, The Netherlands), and sleep states were identified by the resting postures, as described in other birds [6,9,13]. We scored sleep mainly in two behavioural postures when bird had its eyes closed (=front sleep) and when it had tucked its beak on the scapula (=back sleep) [29]. We further described and scored nocturnal sleep in four states (parameters): sleep onset (first sleep bout lasting ≥30 s) and offset (last sleep bout lasting ≥10 s), sleep latency (interval from light off to sleep onset), awakening latency (interval from sleep offset to light on), and sleep frequency (sleep bouts per night). The scores obtained from videographs over 2 days were averaged, and from this mean (±s.e.), bout and duration of sleep both for each hour and total 12 h ‘night' was calculated. Figure 1. Effects on activity and sleep behaviours. Adult female zebra finches were exposed to 12 h light (150 ± 5 lux) coupled with 12 h of absolute darkness (0 lux) or of dim light (dLAN, 5 ± 1 lux). Representative double plotted actogram show a 24 h activity-rest pattern under LD (a) and dLAN (d). Black and white portions denote the active and rest states, respectively. Twenty four hour activity-rest profile in 20 min bins (mean ± s.e.) is also shown for both LD (b) and dLAN (e). Representative sleepogram (24 h distribution of sleep-wake pattern over 2 days) under LD (c) and dLAN (f), with sleep (black) and awake (white) states. (g) illustrates representative images of behavioural postures showing back sleep, front sleep and awake states. Mean (± s.e.) hourly profile of the sleep duration (h), total sleep duration (i), total sleep bouts (j), sleep bout length (k), and sleep and awakening latency relative to lights on and lights off, respectively (l, m) during 12 h of the night period. The asterisk (*) indicates a significant difference, as determined by GLMM (h) or by Student's t-test (i–m). p < 0.05 was considered a statistically significant difference. (*p < 0.05; **p < 0.01).
The activity-rest pattern was recorded throughout the experiment, but for the behavioural assay of sleep birds were videographed over 2 days ending a day before the last day of the experiment. Here, we present data on both activity and sleep behaviours for the same six birds from each condition. Although we monitored behavioural responses of only 6 of 36 birds in this study, we believe that by using an identical experimental procedure, the data are representative of the overall behavioural response pattern of birds in each light condition. As a biomarker of the sleep loss, we measured circulating oxalic acid levels in nine blood samples collected by puncturing the wing-vein in a heparinized capillary tube in the middle of the light period. The handling and blood collection were completed in less than 1 min to avoid any stress-induced effect. Blood was centrifuged for 10 min at 3381 g at 4°C, and plasma was separated and stored at −20°C. Plasma oxalate levels were measured by using a commercially available colorimetric assay kit (Cat. no. K663–100; Biovision Inc., Milpitas, CA, USA), as per the manufacturer's protocol and as detailed in the electronic supplementary material, methods section. These nine blood samples included six birds for which we recorded data on activity and sleep behaviours. At the end of the experiment, birds were sacrificed by decapitation at 4 h intervals beginning just before hour 0 (=light on; n = 6 per time point; hour 0 sampling was completed before light on). Because each bird was separately housed in a light-proof box, it could be taken out for sacrifice without disturbing its neighbour that was still inside its own box; this could be easily verified by activity recordings. Brain from the head kept on ice post-decapitation was carefully removed, snap frozen in dry ice and stored at −80°C until processed for messenger RNA (mRNA) assays. The hypothalamus was excised out as described in Sharma et al. [30], and total RNA was extracted using Tri reagent (AM9738; Ambion), as per the manufacturer's protocol. One microgram of total RNA was treated with RQ1 RNase-free DNase (M6101, Promega, Madison, WI, USA) and reverse transcribed using a Revert Aid first strand cDNA synthesis Kit (Thermoscientific, K1622). At all six times of the day, we measured mRNA expression of six genes of the negative and positive limbs of circadian TTFL (per2, bmal1, reverb-β, ror-α, cry1 and clock), and of eight genes involved in the promotion of sleep (cytokines: tlr4, tnf-α, il-1β and nos; Ca2+-dependent hyperpolarization pathway: camk2, sik3 and nr3a) and the maintenance of awake state (muscarinic cholinergic receptor, achm3) Using gene-specific primers (electronic supplementary material, table S1) and SYBR green chemistry, both target and reference (β-actin) genes were run in duplicates on the Applied Biosystems Step One plus system, and relative mRNA expression levels were calculated by the 2−ΔΔCt method [31], as validated and reported in several publications from our laboratory [11,30]. We used β-actin as a reference (control) gene, which was found to be most stable between three genes that were tested as controls earlier by our laboratory [30]. Further details are given in the electronic supplementary material, methods section. All statistical analyses were performed using GraphPad Prism v. 6.0 and IBM SPSS statistics software v. 20, as appropriate. Student's t-test compared data at one time point between two light conditions. We constructed general linear mixed effect models (GLMM) for behavioural responses. Light condition and time of day were included as fixed effects, and the identity of study subjects was included as a random effect. We also fitted general linear models (GLM) to test the effects of the light condition, time of day, and their interactions on gene expressions. If there was a significant interaction effect, we ran Bonferroni post-test for multiple comparisons. Furthermore, the persistence of a 24 h rhythm in gene expressions was tested by unimodal cosinor regression {y = A+ [B*cos (2*pi (X-C)/24)]}; A, B and C are the mesor (mean value for 24 h expression), the amplitude (maximum change in mRNA expression levels relative to the mesor) and the acrophase (the estimated time of peak mRNA expression) of 24 h (daily) rhythm, respectively. The significance of cosinor regression analysis was calculated using the number of samples, r2 values and predictors—the mesor, amplitude and acrophase (https://www.danielsoper.com/statcalc/calculator.aspx?id=15 [32]). If 24 h variation in mRNA expressions showed a significant daily rhythm, then we used an extra sum of squares F-test to determine the significant difference in the rhythm waveform parameters. For statistical significance, α was set at 0.05. All birds showed a diurnal pattern in the activity-rest behaviour, with activity consolidated during light phase of 12 L : 12 D. However, 24 h distribution of activity showed an overall significant effect of the light condition (F1,720 = 20.102, p < 0.0001) and time of day (F71,720 = 28.446, p < 0.0001), but not of the light condition × time of day interaction (F71,720 = 1.143, p = 0.205; GLMM test). Similarly, sleep was restricted to largely at night (black regions figure 1a); sleep bouts were absent during the daytime (white blank regions, figure 1a) in both LD (dark night) and dLAN (dimly illuminated night) conditions (figure 1). However, we found dLAN-induced alteration in the sleep behaviour. To begin with, the duration of nocturnal sleep bouts showed a significant effect of the light condition (F1,120 = 221.36; p < 0.0001) but not of the time of day (F11,120 = 1.1482; p = 0.147) or the light condition × time of day interaction (F11, 120 = 0.888; p = 0.554) (GLMM test; figure 1d). Particularly, the length (p = 0.0014) and frequency (p = 0.0017) of sleep bouts as well as the total duration of nocturnal sleep (p < 0.0001) were decreased in dLAN birds, compared to LD controls (Student's unpaired t-test; figure 1e–g). Under dLAN, the delayed sleep onsets led to a significant increase in the sleep latency (p = 0.0058), and a significant decrease in the awakening latency (p = 0.0043; Student's unpaired t-test; figure 1h–i). Plasma oxalate concentration faithfully reflected differences in the nocturnal sleep; the levels were significantly decreased indicating sleep debt in dLAN birds, compared to LD controls (p = 0.0286; Student's unpaired t-test; figure 2). Figure 2. Effect of plasma oxalate levels. Mean (±s.e.) plasma oxalate levels in the middle of the day in zebra finches exposed to 12 h light (150 ± 5 lux) coupled with 12 h of absolute darkness (0 lux) or of dim light (dLAN, 5 ± 1 lux). The asterisk (*) indicates a significant difference between the LD and dLAN. p < 0.05 was considered a statistically significant difference.
Figure 3 shows the results. There were significant changes in the level and 24 h rhythm of circadian clock gene expressions (electronic supplementary material, tables S2 and S3). per2 mRNA expression showed a significant effect of the light condition (χ2 = 33.438, p < 0.0001), the time of day (χ2 = 106.920, p < 0.0001) and of the light condition × time of day interaction (χ2 = 23.58, p < 0.0001; GLM test, electronic supplementary material, table S3). In particular, per2 mRNA levels were significantly reduced at hour 4 and hour 8 under dLAN, compared to LD (p < 0.05; Bonferroni's post hoc test). However, reverb-β, cry1 and clock expressions showed a significant effect of the time of day (reverb-β: χ2 = 30.474, p < 0.0001; cry1: χ2 = 56.473, p < 0.0001; clock: χ2 = 14.696, p = 0.012), but not of the light condition or the light condition × time of day interaction (GLM test; electronic supplementary material, table S3). Similarly, bmal1 expression showed a significant effect of the time of day (χ2 = 62.438, p < 0.0001) and light condition × time of day interaction (χ2 = 47.568, p < 0.0001), but not of the light condition (GLM test; electronic supplementary material, table S3). Particularly, as compared to LD, bmal1 mRNA levels were significantly decreased and increased at hour 8 and hour 12, respectively, under dLAN (p < 0.05; Bonferroni's post hoc test). Intriguingly, we found neither the effect of the light condition, nor of the time of day or light condition × time of day interaction on ror-α expression. Most interestingly, cosinor analysis revealed a significant 24 h rhythm in per2, bmal1 and reverb-β expressions in both LD and dLAN, and in cry1, clock and ror-α expressions in LD, but not in dLAN (Fischer's test, electronic supplementary material, table S2). However, we found significant changes in the 24 h rhythm waveform of per2, but not of bmal1 and reverb-β genes. There was a significantly decreased mesor and amplitude, and altered acrophase of per2 mRNA rhythm in birds under dLAN, compared to LD controls (Fischer's test, electronic supplementary material, table S2 and figure 3). Figure 3. Effect on mRNA expression of clock genes. Mean (±s.e.) 24 h mRNA expression of per2 (a), bmal1 (b), reverb-β (c), clock (d), ror-α (e) and cry1 (f) genes measured at 4 h intervals beginning before light on (hour 0 = light on) in the hypothalamus of zebra finches exposed to 12 h light (150 ± 5 lux) coupled with 12 h of absolute darkness (0 lux) or of dim light (dLAN, 5 ± 1 lux). Broken and solid line curves through six time points indicate a significant daily rhythm under LD and dLAN, respectively, as determined by the cosinor analysis. Grey-shaded areas in graphs mark the 12 h night period. The asterisk indicates a significant difference as determined by Bonferroni's post-test after a significant interaction effect indicated by the GLM test (electronic supplementary material, table S3). p < 0.05 was considered statistically significant. Whereas inverted arrow heads indicate the time of peak mRNA expression during the day, the double arrow-headed vertical lines denote the amplitude of daily mRNA oscillations (LD: red; dLAN: blue), as determined by the rhythmometric test (electronic supplementary material, table S2). (Online version in colour.)
Of eight genes that we measured, the cosinor analysis revealed a significant rhythm under LD alone in the mRNA expression of sik3, nr3a and achm3 genes (figure 4; electronic supplementary material, table S3). However, there were changes over 24 h and/or between light conditions in several genes (figure 4). For example, mRNA expression of il-1β showed a significant effect of the light condition (χ2 = 9.069, p = 0.003) and time of day (χ2 = 19.58, p = 0.001; GLM, electronic supplementary material, table S3), but not of the light condition × time of day interaction (figure 4c). Similarly, there was a significant effect of the time of day (tnf-α: χ2 = 50.04, p < 0.0001; camk2: χ2 = 12.33, p = 0.031; achm3: χ2 = 15.79, p = 0.007), but not of the light condition or the light condition × time of day interaction, on tnf-α, camk2 and achm3 mRNA expression (GLM, electronic supplementary material, table S3; figure 4b,e,h). Likewise, tlr4 and nos showed a significant effect of the time of day (tlr4: χ2 = 28.79, p < 0.0001; nos: χ2 = 26.14, p < 0.0001) and light condition × time of day interaction (tlr4: χ2 = 16.79, p = 0.005; nos: χ2 = 13.39, p = 0.02; GLM test, electronic supplementary material, table S3; figure 4a,d), but not of the light condition. sik3 and nr3a expressions also showed a significant effect of the light condition (sik3: χ2 = 9.89, p = 0.002; nr3a: χ2 = 7.39, p = 0.007), time of day (sik3: χ2 = 39.92, p < 0.0001; nr3a: χ2 = 43.39, p < 0.0001) and light condition × time of day interaction (sik3: χ2 = 18.53, p = 0.002; nr3a: χ2 = 49.13, p < 0.0001; GLM, electronic supplementary material, table S3 and figure 4f,g). Particularly, at hour 20, sik3 mRNA levels were significantly reduced in dLAN, compared to LD (p < 0.05; Bonferroni's post hoc test; figure 4f), whereas at hour 12 and hour 4, respectively, nos and nr3a levels were significantly higher under dLAN, when compared with the levels under LD (p < 0.05; Bonferroni's post hoc test; figure 4d,g). Figure 4. Effect on mRNA expression of genes involved in cytokine-induced and Ca2+-dependent gene pathways. Mean (±s.e.) 24 h hypothalamic mRNA expression of tlr4 (a), tnfα (b), il-1β (c), nos (d), camk2 (e), sik3 (f), nr3a (g) and achm3 (h) genes measured at 4 h intervals beginning before light on (hour 0 = light on) in zebra finches exposed to 12 h light (150 ± 5 lux) coupled with 12 h of absolute darkness (0 lux) or of dim light (dLAN, 5 ± 1 lux). The asterisk indicates a significant difference as determined by Bonferroni's post-test after a significant interaction effect indicated by the GLM test (electronic supplementary material, table S3). p < 0.05 was considered a statistically significant difference. Whereas the inverted arrow heads indicate the time of peak mRNA expression during the day, the double arrow-headed vertical lines denote the amplitude of daily mRNA oscillations (LD: red; dLAN: blue), as determined by the rhythmometric test (electronic supplementary material, table S2). (Online version in colour.)
We demonstrate that dLAN at ecologically relevant light intensity levels affected activity and sleep behaviours in zebra finches, similar to those reported in great tits and Indian house crows [6,9]. There was a disruption in the distribution of activity, and significantly delayed sleep onsets and advanced wake ups (hence, a reduction in the duration of nocturnal sleep) in zebra finches under dLAN. In particular, reduced frequency and decreased length of sleep bouts suggested dLAN effects on the consolidation of sleep in zebra finches. Clearly, considering the enhanced sleep latency, zebra finches took a longer time to fall asleep, and spent a greater proportion of the night awake under dLAN, compared to LD. This is consistent with the prolonged daytime activity, and delayed onset of nocturnal sleep in European blackbirds in response to street lights at night [33]. Reduced plasma oxalate levels further evidenced disturbed nocturnal sleep (=sleep debt) in zebra finches in response to dLAN, as has been reported in great tits, rats and humans [8,15]. Because we did not measure plama oxalate levels before the experiment began, we cannot comment whether the large individual variation in levels at the end of the experiment reflected base line differences in oxalate levels or individual responsiveness to dLAN. Nonetheless, the overall significant difference in circulating oxalate levels between the light conditions does suggest the impact of dLAN on sleep. Notably, plasma oxalate levels were decreased in adults but increased in developing great tits in response to the artificial LAN environment [6,8]. It needs to be investigated if effects of dLAN would vary with development stage of the bird. Furthermore, dLAN-induced effect in sleep disruption could be associated with an attenuated nocturnal melatonin peak, because midnight melatonin levels were reduced to almost daytime levels in female zebra finches [11]. Concomitant dLAN-induced negative effects on the sleep-wake pattern and loss of nocturnal melatonin peak secretion have been found in great tits, pigeons and Indian house crows [6,7,9]. We further suggest that dLAN-induced negative sleep effects in female zebra finches involved the endogenous circadian clock, because hypothalamic clock gene oscillations were negatively affected under dLAN. In particular, 24 h changes in cry1, clock and ror-α mRNA levels were arrhythmic, and per2 24 h mRNA rhythm showed an earlier peak and a reduced amplitude under dLAN. This is consistent with an earlier peak expression time of bmal1, per2 and clock, and a delayed peak expression time of cry1 found in tree sparrows (P.montanus) from the urban (artificial LAN), when compared with the rural (no LAN) environment [34]. The abolition of 24 h mRNA rhythm in per2 expression was found correlated with disrupted locomotor activity rhythms and caused sleep loss in mice [35,36]. dLAN also attenuated rhythmic expression of per2 in the hypothalamus, and of bmal1, per2 and cry1 in the liver of nocturnal mice [37]. However, it cannot be ascertained from the present study if dLAN-induced circadian rhythm impairment was causal to sleep disruption or it was the consequence of dLAN-induced sleep disruption in zebra finches. This might represent a self-reinforcing physiological process in which a strongly self-sustained circadian rhythm improves the sleep quality, and sleep homeostasis influences the amplitude (and perhaps phase) of the endogenous circadian rhythm [38]. In general, changes in hypothalamic mRNA levels of genes are consistent with the involvement of multiple hypothalamic neuronal pathways for the promotion of sleep and the maintenance of awake (arousal) state [39]. For instance, the overall effects on tlr4, il-1b and nos mRNA levels are consistent with reduced inhibition of the arousal systems. This, in turn, justifies the awake state promotion under dLAN. Reduced nocturnal sleep and altered tlr4 mRNA levels in zebra finches under dLAN were, in particular, consistent with reduced sleep and enhanced wakefulness in tlr4-deficient mice [18]. With parallel dLAN-induced decrease in il-1β and nos mRNA expressions, we propose a close linkage of the pro-inflammatory cytokines with dLAN-induced sleep fragmentation in zebra finches. Indeed, tlr4 activation induces the synthesis and release of IL-1β and TNF-α, which promote sleep by acting directly on the hypothalamic preoptic neurons, and via nos, cause nitric oxide production [19,20]. The diurnal difference in camk2 expression further suggests the involvement of Ca2+-dependent pathways in dLAN-induced sleep disruption in zebra finches. Changes in camk2, sik3 and nr3a expressions could also suggest dLAN-induced effects on Ca2+-influx and consequently influences sleep in zebra finches, consistent with increased sleep duration in sik3 mutant mice [22]. Concurrently, the lack of a diurnal difference in achm3 mRNA levels under dLAN (opposed to LD) could suggest that the maintenance of awake state was associated with activation of the hypothalamic preoptic M-cholinoreceptors in zebra finches, as reported in pigeons [25]. We caution though that these genes are involved in multiple functions and pathways including sleep, glutamate signalling, calcium influx, thermogenesis and feeding in vertebrates [18,21–24]. dLAN affected the nocturnal sleep and associated hypothalamic molecular pathways in diurnal female zebra finches. Based on changes in mRNA expression of genes comprising the circadian clock circuitry, cytokine and Ca2+-dependent pathways, we suggest an integrated hypothalamic control of sleep-wake state in zebra finches. Thus, by using a diurnal species and ecologically relevant levels of LAN, we closely replicated the prevailing urban night environment, and demonstrated that illuminated nights could negatively impact the molecular underpinnings of hypothalamic sleep-associated pathways in diurnal animals. All procedures were approved and carried out in accordance to guidelines of the Institutional Animal Ethics Committee (IAEC) of the Department of Zoology, University of Delhi, India (Institutional Ethical Approval number: DU/ZOOL/IAEC-R/2018/03). The mRNA sequences with their partial CDS can be accessed using GenBank accession numbers as provided in the electronic supplementary material, table S1. V.K. conceived the idea. V.K. and T.B. designed the study. T.B., I.M. and A.P. performed experiments and carried out sampling. T.B. and I.M. performed assays. T.B., I.M. analysed data, and T.B. and S.K.B. prepared final figures. V.K. and T.B. wrote the manuscript. V.K. and S.K.B. provided all chemical and other laboratory resources. All authors gave final approval for publication. Authors have no competing interests. Financial support from a research grant no. (EMR/2015/002158) by the Science and Engineering Research Board, New Delhi, to V.K. is gratefully acknowledged. T.B., I.M. and A.P. received research fellowships from University Grants Commission of India. Footnotes†[email protected] Electronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.4994600. References
Page 23Animal migration is a ubiquitous global phenomenon, and migratory species play vital ecological roles that are crucial for the functioning of ecosystems [1]. Yet, many migratory animals are rapidly declining due to human activities [2]. Pinpointing the causes and magnitude of these declines is challenging, given the need to understand species’ habitat requirements at multiple sites, often over very long distances. Moreover, many migrants face multiple threats, which further complicate efforts to conserve them. While identifying the magnitude and causes of declines in migratory animals usually requires large-scale monitoring efforts [3–7], insights from site-level studies that consider explicitly the ecology of individual species and site-specific threats are crucial to develop and implement effective conservation strategies to conserve migratory species [8,9]. Migratory shorebirds undertake some of the longest migrations of any migratory animals [10], but they are also declining precipitously worldwide [4,5], especially in the East Asian–Australasian Flyway (EAAF) [11–13]. Shorebird declines along the EAAF have been largely attributed to the loss of tidal flats, which provide critical stopover habitats, in the Yellow Sea region due to coastal development [14–16]. During the 1950s–1980s, more than half of the tidal flats in the Yellow Sea disappeared (−3.0% yr−1); during the 1980s–2000s, the rate of loss slowed down but was still −1.2% yr−1 [17,18]. However, many species along the EAAF have experienced much steeper population declines than the rate of habitat loss. Spoon-billed sandpipers (Calidris pygmaea) declined at an estimated rate of 26% yr−1 during 2002–2009 [19]. Five of the fastest-declining shorebird populations that winter in Australia and New Zealand were all declining at rates ranging between 5.1 and 7.5% yr−1 during the 1990s–2010s [11]. A more in-depth understanding of the nature of shorebird declines along the EAAF may help to conserve not only these species but shorebirds along other flyways as well. The discrepancy between the rates of habitat loss and population decline remains a puzzle. One part of the explanation possibly lies in the fact that not all stopover sites along the EAAF are of equal value to migrating shorebirds [11]; destroying the key stopover sites could have disproportionately detrimental impacts on migrant populations [16,20]. However, where exactly habitat loss occurs within the tidal flats could also matter greatly. Tidal flats are a spatially heterogeneous habitat due to the elevational gradient and the resulting changes in substrate characteristics and benthic invertebrate communities from the high-tide line to the low-tide line [21–24]. Moreover, because of the tidal cycle, only a portion of the tidal flats can be accessed for most of the day at a given site, making the tidal flats temporally heterogeneous. Most of the previous studies of shorebirds' habitat preferences and foraging distributions along other flyways have been conducted during the low-tide periods when all of the tidal flats are freely accessible (e.g. [23,25,26]), while shorebird population monitoring studies are focused on the high-tide periods when birds are packed into a small area and therefore easier to count [20]. Thus, we lack a complete picture as to how shorebirds use the tidal flats, one that takes into account both the spatial and temporal distributions of foraging shorebirds across the entire tidal cycle, especially during ebbing and flooding tides when only a portion of the tidal flats is available to the birds [26]. This gap in our knowledge may impede our ability to develop sound conservation practices for these birds in terms of identifying areas of higher conservation values within the tidal flats. If the portions of the tidal flats that are most important to shorebirds are also the portions that are disproportionately lost due to development, then treating the tidal flats as a homogeneous habitat and averaging the rate of tidal flat loss across the entire range would underestimate the intensity of the threat from coastal development and its potential impact on shorebird populations. Here, we quantified the habitat use patterns and preferences exhibited by migratory shorebird communities at two major stopover sites in the Yellow Sea. We conducted field surveys to simultaneously map the spatial and temporal foraging habitat use of shorebirds throughout tidal cycles. We identify the foraging types based on the species’ use of tidal flats and compare the consistency of these patterns between sites. We also assess the importance of different portions of the tidal flats based on the cumulative time that different species spent foraging on them. In so doing, we identify the upper tidal flat as probably crucially important for the conservation of migratory shorebirds. We conducted fieldwork at two well-known stopover sites for migratory shorebirds in the Yellow Sea region of China: Rudong (32.5 N, 121.2 E) from September to October, 2016, and Nanpu (39.1 N, 118.2 E) from April to May, 2017 (figure 1), coinciding with the peak migration at each site. Rudong is an important site for 18 coastal shorebirds, especially during the southward migration [20,27], and Nanpu supports large populations of at least 12 shorebirds during the northward migration [28,29]. Figure 1. Location of study sites in the Yellow Sea (a) and layout of transects at Nanpu (b) and Rudong (c). At Nanpu (b), each transect contained nine plots of 250 m × 250 m; adjacent transects constituted a transect pair whose data were combined and analysed together. At Rudong (c), each transect contained nine plots of 500 m × 500 m. Plots were numbered 1–9 based on their relative position to the seawall. In (b) and (c), dark grey areas represent land above the high-tide line; stippled areas represent tidal flats and light areas represent the sea beyond the low-tide line. Tidal flat data are from [35].
We set out survey transects on the tidal flats (figure 1) following the design of an earlier study [30]. The transects ran perpendicular to the tidal fronts and covered the entire elevational gradient from the seawall (approx. the average high-spring-tide line) to the low-tide line. The design and layout of transects differed slightly between the two sites because of local conditions: At Rudong, our two transects ran 4.5 km each and were set out so as to avoid areas of high human disturbance, or with numerous tidal channels, or covered by invasive Spartina grass; at Nanpu, four pairs of transects, running 2.25 km each, were set out to increase the coverage of the width of the tidal flat, and the data from each transect pair were analysed together to increase the temporal resolution of data. Along each of these transects, we delineated sets of nine adjacent survey plots, 500 m × 500 m at Rudong and 250 m × 250 m at Nanpu. Each plot was numbered 1 to 9, starting from the plots closest to the seawall, to indicate its relative position along the transects. At various steps of data analysis, we pooled the data from different transects at the same site in order to increase sample size by combining the value or data from plots of the same number. The average distance of each numbered plot to the seawall was calculated for the combined transects, using the distance to the seawall from the midpoint of each plot. In daylight during the spring tides, two to four experienced observers conducted simultaneous surveys of foraging shorebirds in different plots along a given transect. We chose spring tides because they are the periods when the whole tidal flats are flooded and exposed over the course of a full tidal cycle. Each transect was surveyed for 2 to 4 days to ensure that every plot was counted at different times of the tidal cycle, with the cumulative data representing at least two entire tidal cycles pooled for the analysis. For each transect, observers walked parallel to the transects, staying at least 100 m outside the boundaries of each plot to avoid disturbing the birds, and recorded the species and numbers of foraging shorebirds within a plot at the time. The time spent surveying each plot varied depending on the number of shorebirds in it, and the surveyors started counting the next plot as soon as they finished the previous one. Time of exposure/immersion of each plot was also noted to estimate the speed of tidal front movements. Shorebirds were identified to species in most cases. However, three pairs of similar-looking shorebirds were grouped together: great knot (Ca. tenuirostris) and red knot (Ca. canutus, ‘knots’; in Rudong only, as conditions permitted us to identify them to species at Nanpu); Eurasian curlew (Numenius arquata) and Far Eastern curlew (N. madagascariensis, ‘curlews’, both sites); lesser sand plover (Charadrius mongolus) and greater sand plover (Ch. leschenaultii, ‘sand plovers’, both sites). In total, 21 shorebird species (or species groups, referred to as ‘species’ hereafter) were recorded in 827 plot counts. Species recorded in 10 or fewer plot counts at a site were excluded from the analysis, leaving 13 species at Rudong and 13 species at Nanpu. With the data from different transects combined, the average interval (± s.d.) between two counts of the same plot was 15.7 (± 17.3) min, indicating a relatively high degree of temporal resolution in the counting data. For easier interpretation of our analysis and results, we divided transects into three zones, each consisting of three plots, representing the upper (plot 1–3), middle (plot 4–6) and lower (7–9) tidal flats. We recognize that this does not follow the classical delineation of intertidal zones based on the tide levels of different tidal cycles [31]; however, due to the lack of detailed information on the topology of our study sites, we followed previous studies in delineating the zones by dividing each transect into three parts [30,32]. To quantify shorebirds' spatial and temporal distributions, we calculated how the positions of shorebirds’ ‘abundance centroid’ changed with the proportion of tidal flats exposed. A species' ‘abundance centroid’ is the average position of all the individuals during a specific time, and the proportion of tidal flat exposed is a temporal index showing the relative time during a tidal cycle (figure 2). Because a given species may use tidal flats differently depending on the season and site [26], we analysed the counts of species that occur at both Rudong and Nanpu separately, and we will refer to them as two different populations hereafter because of the potential site-specific patterns. Figure 2. Schematic representation of three foraging types exhibited by shorebirds. Boxes in (a, b, c) represent the same stretch of a tidal flat with the seawall (high-tide line) on the left and the low-tide line on the right, when varying proportions of the tidal flats are exposed, representing high, ebbing/flooding and low tides, from top to bottom. The leftmost blue wavy lines represent the location of the tidal front, with the areas to the right of the lines covered by water. Grey dots represent individual foraging shorebirds, and the red triangles represent the positions of the abundance centroid. (d, e, f) The centroid plots show the relationship between the position of the abundance centroid and the proportion of tidal flats exposed. (Online version in colour.)
Based on our field observations and consistent with previous studies [26,30], we proposed three foraging types for shorebirds using the tidal flats: generalists, zone specialists and tide followers (figure 2). Generalists feed whenever and wherever the tidal flats are exposed, thus, on a graph showing the positions of the abundance centroid against the proportion of tidal flat exposed (‘centroid plot’ in short), the slope k will be around 0.5, as the centroid will move towards the middle area when the entire tidal flat is exposed. Zone specialists exhibit a strong preference for a particular part of the tidal flats and remain there even when additional area becomes available. On the centroid plot, the line will be segmented: the slope will increase initially (k1) and then stay unchanged (k2, around 0). Tide followers, as the name suggests, follow the movement of the tidal front and feed in areas that are freshly exposed or about to be submerged. On the centroid plot, the slope k for tide followers will be around 1. The abundance centroids were calculated in three steps. First, we transformed the time when each plot was counted into the position of the tidal front based on an average speed of tidal front movements which we measured during the survey separately for ebbing and flooding tides for each transect. When scaled to 0–1, with 0 representing the seawall and 1 representing the end of a transect (approx. the position of high- and low-tide lines, respectively), this relative position of the tidal front is also the proportion of tidal flats exposed. Next, we combined the counts from different transects at the same site and binned the plot counts into eight tidal periods based on the proportion of tidal flat exposed, in intervals of 0.125. The number, eight, was chosen subjectively as a trade-off between the number of time periods and the number of plot counts in each period. For each plot in a given tidal period, we calculated the average number of foraging individuals of each species. Finally, a species' abundance centroid for a given tidal period at a study site was calculated as the weighted average distance to the seawall, using the distance of each plot to the seawall and the average numbers of individuals of the species in the plot during the tidal period. We normalized the abundance centroids to the same scale as the proportion of tidal flat exposed. To identify the foraging types, we tested the slope k using a linear model starting at (0, 0). We then used the segmented function in R Package segmented [33] to test how the relationship fits a broken line. We chose the following set of criteria to assign each population to a corresponding foraging type based on the results of statistical tests: generalists, 0.45 ≤ k ≤ 0.55 and r2 > 0.90; zone specialists, −0.05 ≤ k2 ≤ 0.05, r2 > 0.90; tide followers: 0.9 ≤ k ≤ 1.0 and r2 > 0.90. This set of criteria was chosen subjectively, and slightly altering them would not change our results and conclusion substantively. Testing whether the specific slope of the relationship falls in the given ranges was done by either testing whether the residuals of the specific linear model overlapped with 0 (for generalists and tide followers), or by overlapping the 95% CI (±1.96 SE) of k2 with the specified range (for zone specialists). Note that these criteria are not mutually exclusive or exhaustive: if a species was assigned to more than one type, we used Bayesian information criterion (BIC) to choose the better fit between the linear and segmented models [34]; if a species was not assigned to any type, we grouped that species as ‘not assigned’. To quantify the importance of different tidal flat zones as foraging habitat to shorebirds, we calculated the cumulative foraging time in bird*minute(s) for each plot across entire tidal cycles. The cumulative foraging time takes into account both how many birds used the area and how long they spent feeding there, providing a comparable and quantitative measurement of habitat importance for within-site comparisons. Assuming the birds were constantly foraging between successive plot counts, for each plot in a transect, we plotted the changes in the number of foraging individuals for each species against the time during the tidal cycle and calculated the area under the curve, i.e. the cumulative foraging time of the species for the plot. To determine the importance of different zones of the tidal flats to the overall foraging by shorebirds, we summed the cumulative foraging time for each species across all transects at the same site. This enables us to show the relative importance of each zone and to simulate the instantaneous impacts of coastal development of the tidal flats on shorebird foraging by calculating the loss of foraging time, assuming that the habitat loss progresses from the seawall to the low-tide line, with no rapid population response or reassortment by the benthic invertebrate fauna upon which the shorebirds feed in the remaining undeveloped portions of the tidal zone. Among the 26 populations recorded, we identified eight generalists, 11 zone specialists, and three tide followers (table 1; figure 3; electronic supplementary material, figure S1). Four populations were not assigned to any type, mostly due to small sample sizes or intermediate patterns. Among the nine species that occur at both Rudong and Nanpu, seven species were assigned a foraging type at both sites, and among them, only two species, the Terek sandpiper (Xenus cinereus) and bar-tailed godwit (Limosa lapponica), exhibited the same foraging type at both sites. Figure 3. Distribution and centroid plots of four representative shorebird populations. We show here the representative populations of the three foraging types, and one population that was not assigned. In the spatiotemporal distribution plots (a, b, c, d), the number of foraging individuals recorded in a plot count is represented by the size of the solid blue circles, and the red open circles represent counts with 0 individuals. In the centroid plots (e, f, g, h), solid circles represent the positions of abundance centroid calculated for each tidal period along the tidal flats (0: seawall/high-tide line; 1: low-tide line). The solid line is the best-fitting line for the centroids passing through (0, 0), with 95% confidence intervals shown in grey. The dotted reference line has a slope of 0.5 (generalist) or 1 (tide follower). The dashed line represents the predicted broken-line relationship (zone specialists). All analysed populations are in electronic supplementary material, figure S1. (Online version in colour.)
For most shorebirds among the 26 populations, the majority of their foraging time was spent in the plots closer to the seawall, shown by the cumulative foraging time along the tidal flats (figure 4; electronic supplementary material, figure S2). This is especially the case among generalists and zone specialists; tide followers predictably divided their foraging time roughly equally across the different plots or zones. For example, the proportions of cumulative foraging time spent by a generalist, the Kentish plover at Rudong, on the upper, middle and lower tidal flats, respectively, were 74.5%, 17.5% and 8.0%; for a zone specialist, the Terek sandpiper at Nanpu, these values were 99.0%, 1.0% and 0% and for a tide follower, the sanderling at Nanpu, they were 17.7%, 46.8% and 35.5%. Figure 4. (a–d) Cumulative foraging time along the tidal flat for four representative shorebird populations. The cumulative foraging time of the same species in figure 3 is shown as the proportion of total cumulative foraging time for a given population at a given site with transects summed. Solid lines show the distribution of cumulative foraging time of each plot along the tidal flat. Dashed lines show the reduction in total cumulative foraging time that would result from the hypothetical loss of tidal flats starting from the seawall and progressing seaward. (Online version in colour.)
We aggregated the cumulative foraging time and plotted the loss of total foraging opportunity that would occur in response to a hypothetical loss of tidal flats that progresses from the seawall outward into the ocean to simulate the instantaneous effect of coastal development (dashed lines in figure 4; electronic supplementary material, figure S2 for individual species; solid lines in figure 5 for all species in a given foraging type). Losing the upper third of the current tidal flats would lead to a reduction in cumulative foraging time of, on average (± s.d.), 77.6% (± 13.6%) for generalists, 81.0% (± 20.9%) for zone specialists, 23.5% (± 17.0%) for tide followers, and 75.0% (± 32.3%) for species not assigned to any foraging types. Figure 5. (a–d) The importance of upper tidal flats to shorebirds of each foraging type. Solid lines in each plot show the reduction in total cumulative foraging time assuming habitat loss of the tidal flats starting from the seawall for all populations within the specified foraging type, in which the thicker line is the representative population in figure 4. The black dashed line shows the average proportion of decline of the type, and the dotted line is a reference line showing a proportional reduction when different portions of the tidal flat are of equal importance. (Online version in colour.)
The conservation of migratory species requires an understanding of both their ecology at a fine spatial scale and the site-specific threats they face. By quantifying the spatial and temporal patterns in shorebird foraging distributions throughout the tidal cycle at two critical stopover sites in the Yellow Sea, we found that 17 migratory shorebirds exhibit substantial interspecific and even site-specific differences in their use of different portions of the tidal flats (table 1; figure 3; electronic supplementary material, figure S1). These populations can be mostly grouped into three foraging types: generalists, zone specialists and tide followers. Notwithstanding these differences in foraging behaviour, however, the upper tidal flat zone provides the majority of the cumulative foraging time of both the generalists and the zone specialists, far greater than the upper tidal zone's proportional area (figures 4 and 5; electronic supplementary material, figure S2). Because coastal development projects typically start near the high-tide line and proceed outward toward the sea [35,36], the upper tidal flats are also more prone to development than are the lower tidal flats, which may help to explain why shorebird populations along the EAAF have declined much faster than the overall rate of stopover habitat loss. Our work highlights the need to conserve as much of the upper tidal flats as possible within important stopover sites in order to protect today's diminished populations of migratory shorebirds. Our study also demonstrates the value of understanding the detailed patterns of habitat usage by migratory species throughout their journeys in order to properly conserve them. By taking into account the temporal changes in habitat availability, we show that most shorebird species at the two stopover sites fall into the three foraging types. Previous studies along other flyways about the foraging distribution of shorebird communities and their relationship with food availability have focused almost entirely on the spatial pattern during low tides [23,25], thereby overlooking the temporal changes in habitat (and thus food) availability and their effect on shorebirds' foraging behaviour during ebbing and flooding tides (see [26,30,37,38] for studies that considered tidal movements). Our analysis also shows that the same species may display different foraging behaviours at different stopover sites (table 1; figure 3). Of the seven species that occurred at and could be assigned to a foraging type at both sites, only two exhibited the same type, suggesting that these foraging behaviours may not be a species-specific behavioural property but, as suggested by earlier studies, may change in response to local environmental conditions [38–41], or differ between sexes or subspecies that constitute the local populations [42,43]. This result, while statistically robust for some species, needs to be validated by further studies. To simultaneously quantify the spatial and temporal aspects of shorebirds’ foraging distributions, our statistical tests rely on a single-derived and consolidated property (the position of shorebirds' abundance centroid in relation to the proportion of tidal flats exposed), which limits our ability to identify and account for finer patterns and changes in foraging distributions (e.g. red knot at Nanpu, in electronic supplementary material, figure S1). Our approach is expected to be especially poor at distinguishing between a generalist and a zone specialist, because a zone specialist will gradually turn into a generalist if its preferred foraging zone covers an increasingly large portion of the tidal flats. This was evident in our analysis when we had to rely on BIC to choose the better model describing the distributions of two populations that fell into both the generalist and zone specialist categories. There is clearly a need for more sophisticated statistical methods and more comprehensive field studies focused on the foraging behaviours of shorebirds using tidal flat habitat. We show that most shorebird populations at the two stopover sites spend disproportionately large amounts of time foraging on the upper tidal flats, particularly the generalists and zone specialists (figure 5), a combined result of the upper zone's longer exposure time and the shorebirds' preference for it. It may seem counterintuitive that the upper tidal flats are almost as important for the generalists (77.6% ± 13.6%) as they are for the specialists (81.0% ± 20.9%), but this reflects the relatively minor contributions of the middle and lower tidal flats to the cumulative foraging time of the generalists, a consequence of the limited exposure time relative to the upper tidal zone. Unfortunately, the upper tidal flats are also the part of the habitat that is most prone to coastal development [30,35,36]. Decreases in the amount of available foraging habitat and foraging time could lower shorebirds’ survival rates and/or breeding success, directly or via carryover effects [8,44,45]. By simulating progressive seaward development of the tidal flats, we predicted that the loss of upper tidal flats may cause substantial and disproportionately severe reductions in overall foraging opportunities (figures 4 and 5), assuming the benthic invertebrate species upon which the shorebirds feed do not reassort in a way that partially compensates for the loss of foraging opportunities. Note that our results were generated from surveys conducted only during the spring tides, when the entire tidal flats are covered and then exposed during a tidal cycle, leading to more uniform distributions of foraging time among different zones. During a neap tide, the upper tidal zone would be exposed even longer relative to the lower tidal zone, suggesting that the overall importance of upper tidal flats to the foraging shorebirds could be much higher than we showed in the current study. We suggest that the disproportionate importance of the upper tidal flats to EAAF shorebirds, combined with their heightened historical and current vulnerability to coastal development, may help to explain the discrepancy between the overall rates of coastal habitat loss (−1.2% yr−1 [17]) and the rate of population decline in EAAF shorebirds (up to −26% yr−1 in a similar period [11,19]). Thus, analyses based on the average rate of habitat loss may have significantly underestimated the severity of the threat that coastal development poses to EAAF shorebird populations. However, we could not perform any correlational analysis to confirm this hypothesis, as we lack detailed information on the changes in the total area of the different tidal flat zones in the Yellow Sea region, especially with respect to exactly how much more of the upper tidal flats was lost relative to the lower tidal flats, apart from the fact that most of the development has proceeded from the high-tide line outward to the sea, impacting the upper tidal flats most severely [17,35,36,46]. We recognize that our measure of tidal flat importance looked only at the number of foraging individuals and their foraging time, but not their intake rates. However, other studies have indicated that the higher intake rates often achieved at lower tidal flats are insufficient to compensate for the shorter exposure times [30,47] and that shorebirds select areas where the tidal flats are exposed for longer periods [48]. Another limitation of our study is that our simulation of upper tidal development takes into account only the instantaneous impact of habitat loss and assumes no reassortment of the benthic invertebrate species upon which the birds feed or other changes in local conditions. Coastal development in the long term, however, not only leads to direct habitat loss, but also alters the remaining habitat by changing its hydrodynamic and deposition patterns in ways that can lead to erosion or siltation [35], which, in turn, affect the substrate structure, size, physiochemical conditions and benthic invertebrate faunas of stopover sites [21,49]. A comprehensive and long-term monitoring programme focusing on the changes in both benthic fauna and shorebird distributions is needed to provide a mechanistic understanding on how coastal development affects local shorebirds both instantaneously and in the long term, as well as whether and how fast the tidal flats may recover their sizes and benthic invertebrate fauna. As the most valuable yet most vulnerable parts of the tidal flat for shorebirds, the upper tidal zones should be the focus of conservation actions within the Yellow Sea region, through a combination of protecting them from further development and improving the quality of existing habitat. The planning of future development projects in this region should entail careful consideration of any activities that disproportionately affect the upper tidal flats and/or areas providing major foraging opportunities identified by local studies following our approach. For threatened species with very small population sizes (e.g. spoon-billed sandpipers), tracking the movement of individuals throughout the tidal cycle using a combination of direct observations and telemetry may be needed to understand their habitat preferences [9], in addition to our population-level approach. We encourage researchers to conduct similar studies on the spatiotemporal distribution of foraging shorebirds elsewhere along the EAAF and along other flyways to better understand how shorebird species are using the tidal flats throughout their annual migration and to determine if the upper tidal flats are of disproportionate importance to these birds elsewhere. Development of tidal flats is not limited to the Yellow Sea. A recent study showed extensive losses of tidal flats around the world: more than 15%, or 20 000 km2 worldwide between 1984 and 2016 [35]. Such development, combined with sea-level rises triggered by global climate change [50], simultaneously threatens the tidal flats and many of the world's shorebirds. Absent a concerted, international effort to protect the intertidal stopover sites these birds depend upon, the world stands to lose some of its most remarkable long-distance migrants. The field survey protocol was approved by the Institutional Animal Care and Use Committee of Princeton University. Data and R code are available from the Dryad Digital Repository: https://doi.org/10.5061/dryad.gmsbcc2jb [51]. T.M. perceived and designed the study, collected and analysed the data, and drafted the manuscript. D.S.W. codesigned the study, provided critical input for analysis and extensively revised the manuscript. All authors gave final approval for publication. This work was supported by the High Meadows Foundation. We thank U. Srinivasan, F. Hua, H. Peng and the Drongos for useful discussions throughout the study. We thank Y. Wang, X. Shi, J. Zhang, C. Yu, Y. Zhong, X. Chen, W. Liu, J. Li, J. Chen, F. Yan, C. Liu, J. Loghry, T. Zhao, L. Zhang, J. Li, W. Lei, Y.C. Chan, A. Boyle, C. Hassell, Z. Zhang and Z. Ma for their assistance and logistical support in the field. We thank X. Ren and X. Shang for preparing figures 1 and 2. We also thank the editors and reviewers who greatly improved the manuscript. FootnotesElectronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.4977719. References
Page 24Body size is a key functional trait, dictating energy demand, prey preferences, and the overall ecological role of animals [1–6]. The accurate prediction of future ecological changes may thus depend heavily on a mechanistic understanding of how key functional traits like body size are affected by environmental changes [7–9]. Today, the most common environmental change is an increase in average local temperatures [10]. In the majority of ectotherms, higher rearing temperatures are associated with faster growth rates and smaller adult body sizes [11–14]. As such, the plastic ‘temperature-size rule’ of fast growth but reduced adult body size has become incorporated into models predicting the population and community outcomes of warming [15–18]. While this plastic response is often assumed to be adaptive [19,20], the role of evolution in modifying this response is unknown over the short time scales relevant to current warming. Evolutionary responses to temperature are poorly understood because confounding factors are common along thermal gradients in nature. Latitudinal and elevational temperature gradients are commonly confounded with other putative selective agents like precipitation, resource availability, biogeography and seasonality [21–23]. Experimental evolution studies can overcome this issue by isolating the effects of temperature as an agent of selection. However, these experiments have been restricted to simplified, controlled laboratory environments (but see [24–26]), and they have been limited to testing for evolutionary responses to temperature in smaller-sized taxa (e.g. plankton, Drosophila, etc. [27–33]). Thus, surprisingly little is known of how warming per se may cause the evolution of ecologically important traits like body size and growth rates for larger taxa in nature. Animals in aquatic environments display the strongest and most consistent temperature–size responses [13,34], indicating there may be a common adaptive explanation for this pattern. Warmer water temperatures are likely to increase oxygen limitation and may also increase resource limitation [19,35]. Large individuals have higher overall metabolic demands [5], so they may be most challenged by this decrease in availability. If large individuals are more strongly challenged by warming, life-history theory can provide simple explanations for the evolution of reduced body size. For example, if oxygen or resource limitation increases mortality rates for large size individuals, then life-history theory predicts the evolution of earlier and greater reproduction at a smaller size [36]. Relatedly, if oxygen or resource limitation stresses large size individuals, reducing the fecundity advantage of large size, life-history theory also predicts the evolution of earlier and greater reproduction at a smaller size [36,37]. Thus, if warming alters mortality or fecundity selection in a manner that disfavours large individuals, then evolution may contribute to the expression of earlier and greater reproduction at a smaller size. Adaptive or not, the temperature–size rule describes the pattern of warming-induced size reduction at a given life stage (e.g. parturition, maturity) via plasticity, but warming could also affect body sizes through the evolution of body growth rates. For example, warming could cause the evolution of reduced somatic growth rates after maturity as a simple by-product of greater reproductive investment [38–40]. Warming may also cause the evolution of reduced growth rates before maturity, counteracting plastic growth acceleration at that life stage. For example, populations of a variety of fish taxa show evolved increases in juvenile growth rates at higher latitudes compared with populations of the same species from lower latitudes, a pattern that counteracts plasticity and promotes growth rate similarity across environments [41–47]. However, this ‘countergradient’ pattern in growth is thought to emerge along latitudinal gradients due to variation in the length of the growing season, with higher-latitude populations evolving fast growth to overcome the shorter length of the growing season [23,48]. Therefore, it is less clear that increased temperature per se will drive the evolution of reduced growth rates under current warming. Here, we sought to test the hypothesis that increased temperature disfavours large body size, causing the rapid evolution of reduced somatic growth rates and a shift in allocation towards greater reproduction at a smaller body size. We used populations of western mosquitofish (Gambusia affinis) recently established across a unique set of geothermal springs. Geothermal springs can provide useful thermal gradients that break confounding patterns found along other natural thermal gradients such as latitude and altitude [49]. Past work in this study system suggests that mortality rates may be higher for larger individuals at warmer temperatures. Populations from warmer sites tend to have smaller body size distributions [50], despite that individual growth rates tend to increase over this range in temperatures [51,52]. Moreover, field routine metabolic rate measurements suggest that the mass-specific metabolic advantage of large size is lost at higher temperatures; metabolic scaling coefficients shift from the metabolic theory expectation of approximately 0.75 at cooler temperatures towards approximately 1 at higher temperatures [50]. We tested four predictions about selection and evolution of body size and growth in warmed environments. First, we tested whether warming alters fecundity selection to favour greater reproduction at a smaller size. To do so, we used a trait survey of wild-caught female fish, to test whether higher temperatures were associated with a decrease in the fecundity advantage of large body size. Second, using first laboratory generation (F1) adult females reared in a common environment, we tested the prediction that warmer-source populations have recently evolved an increase in reproductive effort at small sizes. Third, we measured the embryo size of these mothers to test the prediction that evolution contributes to a reduction in offspring size under warming. Fourth, we used second laboratory generation (F2) juveniles reared across five temperatures to test the prediction that warmer-source temperatures were associated with the recent evolution of reduced somatic growth rates. We expected growth rates to increase with rearing temperature due to plasticity alone, so this pattern of evolution opposing plasticity would demonstrate the recent evolution of countergradient variation in growth. The western mosquitofish is a small (less than 6 cm), sexually dimorphic, livebearing fish that was introduced across the globe throughout the twentieth century [52] (figure 1a). Male mosquitofish virtually cease growth at maturity, while female mosquitofish, like both sexes in many fishes, exhibit indeterminate growth [53]. Mosquitofish exhibit a thermal niche common of temperate ectotherm species; they can tolerate temperatures from near-zero to approximately 40°C but require warmer minimum temperatures for reproduction (approx. 16°C [52]). Here, we study populations spanning most of this reproductive thermal range. Figure 1. Features of the study system. (a) One of the seven geothermal spring ponds where mosquitofish were collected for this study (‘AW,’ 23.7°C). Inlayed is a photograph of mosquitofish in this pond. (b) A map of the geothermal spring sites (Inyo and Mono counties, California). Coloured points are sampling sites, and colours correspond to the temperature gradient. Black points are landmarks. (c) Temperature profile of each geothermal spring, logged at 15 min intervals over spring 2014. For reference, the daily average air temperature (Bishop Airport, Bishop, CA) is plotted in black. The temperature logger at the warmest site (‘LHC’, 33.3°C) failed at midday on 25 February. (Online version in colour.)
Mosquitofish were introduced into a single site in California (CA) in 1922 from 1 to 2 sources in Texas [54]. Mosquitofish were then spread widely within the state but documentation of their timeline and introduction pathway into specific sites is rare. Today, mosquitofish occupy geothermal springs in Inyo and Mono counties (figure 1b). The focal springs studied here are unique in that they are dammed near the spring source (electronic supplementary material, table S1). Consequently, the populations of mosquitofish in these springs comprise individuals experiencing a consistent and highly constrained thermal regime, with virtually no gene flow from other environments. Each spring has a different mean temperature, and the springs exist in close proximity, relatively evenly placed along a gradient from 18.8 to 33.3°C (figure 1; electronic supplementary material, table S1). We monitored water temperature at the focal sites for several months over a period of significant change in air temperatures to confirm that these sites were thermally stable (figure 1c). For trait analyses on wild-caught fish, we collected female mosquitofish using seine and hand nets. We sampled in spring and in summer to coarsely assess seasonality effects (see electronic supplementary material, table S2 for sample sizes and dates). Fish were euthanized with an anaesthetic overdose, immediately preserved in 95% ethanol, and later stored in 80% ethanol. For common-garden rearing, we collected wild fish from six of the seven populations on 18 February 2018 and transported them to the University of California (UC) Santa Cruz. One site—‘LAW’ (24.8°C, electronic supplementary material, table S1)—had been invaded by several predatory largemouth bass (Micropterus salmoides) in the period between field collections and before collections for common rearing (D.C.F. 2015, personal observation), so we did not use that site for springtime sampling or common rearing. Other sites do not contain large piscivorous fishes (electronic supplementary material, table S1). Fish rearing for the wild-caught and first laboratory-born generations (F1) took place in an environment-controlled greenhouse at UC Santa Cruz (electronic supplementary material, figure S1). We introduced wild-caught adult fish to 568 l tanks (n = 6 tanks, 1 per population; electronic supplementary material, figure S1) at a density of 40 ± 5 individuals per tank. Tanks were identical and tank assignment was randomized. Tank water was off-gassed Santa Cruz city water. Each tank was heated to 26°C with a 500-watt submersible heater and water temperature was homogenized in each tank by continuous, vigorous bubbling with an air pump. Photoperiod was set to 14 : 10 h daylight : dark using full spectrum lighting. Fish were fed ground Tetramin (Tetra Holding, Blacksburg, VA, USA) flake food in morning and evening and Frystartr (Skretting Inc, Stavanger, Norway) midday. Water quality was maintained through siphoning of waste and 50% water changes twice weekly. Newborn offspring were collected on floating fry retention devices that reduce cannibalism by adults (electronic supplementary material, figure S1). Experimental offspring were collected twice daily and retained for F1 rearing starting 18 March 2018. We waited this one-month period from adult collection to fry collection to ensure the offspring we collected were not directly exposed to their parent's natal thermal environment during early internal development. The interbrood interval of mosquitofish is about 20 days at 30°C [51]. All newborn fish from the same population born on the same day were reared together in a ‘fry basket’ hung in 57 l tanks in the same room and also set to 26°C (electronic supplementary material, figure S1). By 15 April 2018 we had collected at least 90 F1 fish from each population, representing estimated genetic contributions of at least 12 females per population, but probably many more (electronic supplementary material, table S3). At that point, F0 fish were euthanized, their tanks were drained, cleaned and reset, and F1 fish were introduced. F1 fish were haphazardly reduced in density to 72 ± 6 individuals for each tank. Additional F1 fish not transferred to 568 l tanks were reared in the 57 l tanks until 4 July 2018, when they were euthanized and preserved as above. On 16 June 2018, we began collection of second laboratory generation (F2) offspring as for F1 fish above. For F2 trait assays, we collected up to 10 individuals born per population per day. We continued to collect F2 fish until December 2018, at which point F1 fish were euthanized and preserved as above. We dissected wild and F1 female mosquitofish to obtain four life-history traits related to reproduction and offspring size: reproductive allocation (gonadosomatic index, hereafter ‘GSI’, calculated as gonad weight ÷ total weight), fecundity (number of embryos), mean embryo diameter and mean embryo mass (F1 fish only). We sought to obtain trait data from females across a similar range in body size from each site. We also measured embryo stage as a potential covariate affecting these traits, using a modified protocol from [55] in which one of six development stages was assigned to each brood. Our modified trait measurement protocol is provided in electronic supplementary material, appendix S1. We tested for effects of source temperature and maternal body length on the focal adult traits. We excluded non-gravid females, which were rare, from the dataset. We assigned each fish a source temperature which was the average from the time series of the population's source temperature (figure 1; electronic supplementary material, table S1). We used generalized linear models for each trait by dataset (wild in spring, wild in summer, F1) combination. We sought to remove (control for) the independent covariate effect of embryo stage, which we expected could influence life-history traits (e.g. since embryo diameter increased with embryo stage; electronic supplementary material, figure S4). We started with the full model specification: trait ∼ maternal length × source temperature+embryo stage. When the interaction was not significant, we removed that effect and re-ran the model. For GSI we used a Gaussian error distribution, which performed well because there were few values near 0% or 100%. For fecundity, we used a quasi-Poisson error distribution because data were counts and overdispersed. We did not include the covariate effect of embryo stage in models predicting fecundity, because preliminary analyses demonstrated that effect was non-significant (all p > 0.07) in each case, and past work shows partial atresia (a reduction in embryo number through embryonic development) is unsupported in mosquitofish [56]. For embryo diameter and embryo mass, we used a Gaussian error distribution, which performed well after log10 transformation of those response variables. We constructed these models in the R environment using the glm() function [57]. To obtain model coefficients and associated p-values, we used the summary() function. To obtain model R2 values for the fecundity model, we used the package ‘rsq’ [58]. To approximately visualize models for GSI and log10 embryo size without the influence of embryo stage (the covariate), we obtained residuals from the OLS regression trait ∼ embryo stage, and plotted predictions for the model residuals ∼ maternal length × source temperature. To assay juvenile growth rates, we reared newborn F2 fish across five treatment temperatures in two controlled environment rooms (TriMark R.W. Smith, San Diego, CA, USA) at UC Santa Cruz (electronic supplementary material, figure S2). The air temperature in one room was set to 23°C, and included tanks with treatment water temperatures 23, 29 and 32°C. The air temperature in the other room was set to 19°C, and included tanks with treatment water temperatures 19 and 26°C. The environmental settings in these two rooms were otherwise set to be identical, including a photoperiod set to 15 : 9 h daylight:dark. Tanks were 100 l plastic tubs (91 × 61 × 20 cm3) filled with offgassed Santa Cruz city water. There were five replicate tanks per treatment temperature. In each room, tanks were randomly assigned a treatment temperature. Tanks with treatment temperatures above set air temperatures were warmed with submersible aquarium heaters. Water in all tanks was continuously homogenized with submersible water pumps (150 l per hour) to prevent thermal gradients within tanks. Tank water temperatures were monitored daily. We maintained water quality through siphoning of waste and 90% water changes biweekly. Fish were reared individually in cylindrical mesh containers with a Petri dish bottom and an open top (250 µm mesh, 7 cm diameter, 20 cm height). Fish containers were sunk into the water tanks, with the open top of container several centimetres above the water line, to prevent fish escape. Fish of a given source population were assigned temperature treatments sequentially as they were born, so that each population had approximately equal representation across the rearing temperatures. Fish of a given treatment temperature were then assigned one of the five replicate tubs sequentially as they were born, such that fish density differences among tubs were minimized through time. Fish were fed an excess of Frystartr food (Skretting, Stavanger, Norway) thrice daily. Growth was measured as the difference in total length at age 0 and at age 15 days. Lengths were measured from top-down photos taken with a scale bar (electronic supplementary material, figure S3), and analysed in ImageJ software [59]. We tested for effects of source temperature on newborn size and on juvenile growth rates. For newborn size, we used the OLS regression: log10(length) ∼ source temperature. For juvenile growth rates, we used the OLS regression growth∼source temperature × (rearing temperature)2. We included the second-order polynomial to allow for curvature in the effect of rearing temperature, as preliminary plots showed a curved pattern in growth across rearing temperatures. To do so, we used the poly() function in R, which creates orthogonal first- and second-order terms to allow interpretation of the significance of these coefficients separately. The interaction term was non-significant (p = 0.332), so we dropped that term from the model. Models were constructed using the lm() function. To obtain model coefficients and associated p-values, we used the summary() function. Diagnostics of the model predicting growth rates indicated possible violations of the homoscedasticity and normality assumptions (due to leptokurtosis). To evaluate whether this issue caused substantial problems for parameter estimates and statistical significance, we compared the model output with three ‘robust’ regression methods that deal with these issues. We did not observe substantive differences in the model output (electronic supplementary material, table S4), so here we report output from the simpler OLS regression. Finally, to test for differences in survival to age 15 days across source populations and rearing temperatures, we used a generalized linear model with a binomial error distribution. We used the model specification survival ∼ source temperature × rearing temperature, with rearing temperature treated as a factor. Data from this manuscript are available on Dryad [60]. From field surveys in springtime, there was no support for an effect of site temperature on GSI or fecundity, though fecundity did increase with body size (figure 2a,d and table 1). However, by summer, warmer-source populations showed a weaker increase in GSI and fecundity with increasing body size than did cooler-source populations (figure 2b,e), indicating that larger individuals performed relatively poorly at higher temperatures. Moreover, summertime samples indicated that individuals from warmer sites had higher fecundity at small sizes than did similarly sized fish from cooler sites (figure 2e). After common rearing, warmer-source populations expressed relatively high GSI and fecundity across all body sizes (figure 2c,f), supporting that recent evolution has led to an increase in overall reproductive effort in warmer-source populations. Figure 2. Effects of mosquitofish body length and source temperature on the focal life-history traits. To visualize these effects without the covariate effect of embryo stage (for GSI and embryo diameter only), we used residuals from the model trait ∼ embryo stage, and then plotted predictions for the model trait residuals ∼ maternal fish length × temperature. Interaction terms were always included for this graph, even if they were non-significant and removed for the final analysis in the main text (table 1). Points and prediction lines are coloured as in figure 1, with the 18.8°C source population labelled the darkest blue and the 33.3°C source population labelled the darkest red. Significant effects (p < 0.001***, 0.01** 0.05*) are included in the bottom-right corner of each panel (see table 1 for full statistical output). (Online version in colour.)
In the wild, females from warmer-source populations had relatively small embryos, but common rearing showed weak support that this was caused by recent evolutionary divergence among populations. In both spring and summer, warmer-source populations had significantly smaller embryos (measured as diameter), and embryo size was unrelated to maternal size (figure 2g,h). After common rearing, we measured three metrics of offspring size. Analysis of embryo diameter suggested a weak interaction between maternal size and source temperature, such that cooler-source populations showed a slight increase in embryo size with increasing maternal size. This effect was reduced with increasing source temperature, as warmer-source populations showed weak to no evidence of increased embryo diameter with increasing maternal size (figure 2i). Analysis of mean embryo mass showed no interaction or effect of source temperature, though embryo mass did increase with maternal size. Finally, analysis of F2 newborns showed no evidence that source temperature affected newborn length (p = 0.292; electronic supplementary material, figure S5). Thus, across these three metrics, there was weak evidence that source temperature caused substantial evolution of offspring size, and there was also weak evidence that maternal size substantially influences offspring size in this species. Across all rearing temperatures, second laboratory generation newborns from warmer-source populations exhibited slower growth rates than did newborn fish from cooler-source populations (figure 3a). We used model predictions for the coolest-source population reared at all temperatures and compared it with each source population theoretically reared at exactly its own source temperature. This exercise showed that ignoring evolution causes an overestimation of the acceleration of growth under warming (figure 3b). Moreover, with increasing magnitudes of warming, there is an increase in the importance of evolution as a proportion of growth (figure 3b). Finally, survival analysis on these F2 fish over their first 15 days of life showed approximately 75% overall survival and indicated no significant differences among source populations or rearing temperatures (all p > 0.401; electronic supplementary material, figure S6). Consequently, there is little evidence that selection during the second generation of rearing could have led to (a reduction of) observed differences in growth. Figure 3. (a) Effects of source temperature on juvenile growth rates across the five rearing temperatures. Points and prediction lines are coloured as in figure 1, with the 18.8°C source population labelled the darkest blue and the 33.3°C source population labelled the darkest red. Points are jittered along the x-axis. Significant effects (p < 0.001***, 0.01** 0.05*) are highlighted in the top-left corner. (b) Predicted temperature dependence of growth rate for the 18.8°C source population (plasticity only) or for each source population reared at its own source temperature (plasticity + evolution). Shown are the percentage reductions in growth caused by evolution at each temperature relative to the 18.8°C population. Error bars represent 95% confidence intervals of the prediction. Note the difference in ranges on y-axes of each panel. (Online version in colour.)
We used recently established populations of mosquitofish (Gambusia affinis) across a unique geothermal temperature gradient from 19–33°C to test for effects of temperature on the recent evolution of size-related traits. Trait surveys of wild-caught fish indicated that higher temperatures reduce the benefit of large size. Specifically, warmer-source populations showed a weaker increase in GSI and fecundity with increasing body size compared to cooler-source populations. Higher temperatures were also associated with a substantial reduction in embryo size in the wild. After common-environment rearing to reveal recently evolved trait differences, warmer-source populations showed little difference in embryo or offspring size relative to cooler-source populations, indicating that the warming-induced reductions in offspring size observed in the wild may be caused by plasticity. However, after common rearing, warmer-source populations did exhibit a relative increase in reproductive effort and fecundity at small sizes and a decrease in juvenile growth rates. Altogether, these data are consistent with the hypothesis that warming disfavours large body size, leading to the evolution of increased reproduction at small sizes and to slowed somatic growth rates. Our results support the notion that natural selection at warmer temperatures favours reduced size, and therefore that the plastic temperature–size rule may be adaptive. Specifically, our results show that warming alters fecundity selection, reducing the fecundity advantage of large body size. A recent study showed a similar reduction in the fecundity advantage of large size in freshwater snails, and a simple life-history model suggested that this would favour the evolution of the temperature–size rule [37]. Here, in alignment with this life-history model, we show that populations from warmer sources have indeed recently evolved an increase in reproductive effort and fecundity at small body sizes. Interestingly, common-reared warmer-source populations also showed relatively high reproductive potential at large sizes, indicating that warmer-source populations are predisposed to greater reproduction throughout life. However, in nature, warmer-source populations were not able to sustain this high reproduction to later ages and larger sizes, probably resulting from the stressors associated with living at higher temperatures and under natural conditions. Although our data suggest that altered fecundity selection may contribute to the evolution of earlier and greater reproduction at a smaller size, our past work on mosquitofish populations in these geothermal springs also indicates that other forms of natural selection may be altered at warmer temperatures. In the wild, average female body sizes are smaller at warmer temperatures [50]. Female mosquitofish grow continuously throughout life, and their growth rates generally increase with temperature [51,52]. Therefore, it appears that mortality rates are higher for larger individuals in warmed systems, which life-history theory predicts can also lead to the evolution of the greater reproduction at a smaller size that we observed here [36]. Importantly, increased oxygen and resource stresses have been proposed as general mechanisms favouring reduced size under warming in aquatic environments [19]. Our data suggest multiple types of selection (mortality, fecundity) resulting from stressors in aquatic environments may contribute to similar patterns in body size evolution. These results may help to explain the widespread nature of the temperature–size rule in aquatic taxa [13], and may explain why terrestrial organisms show less consistent patterns of selection [61]. Looking forward, a more precise understanding of the drivers underlying warming-induced evolution could be achieved by parsing the relative contributions of altered mortality versus fecundity selection for different species and in different environments. In our study, common-reared warmer-source populations displayed slower juvenile growth rates across a wide range of rearing temperatures. This countergradient variation in growth rates has been found along latitudinal temperature gradients in many other fish species [23]. In those cases, it had not been clear whether temperature was the driver of countergradient variation in growth rates, because factors like seasonality of light and resource availability also vary systematically along those thermal gradients. In our study system, the thermal gradient does not appear to be strongly confounded with at least two metrics of basal resource availability—nutrients and chlorophyll a (electronic supplementary material, table S1). More importantly, a large component of the diet of mosquitofish in these springs comes from allochthonous, or outside, sources (i.e. terrestrial insects), effectively decoupling local resource production from resource availability (E.R.M. 2016, personal observation). In addition, the seasonality of photoperiod was not likely to covary with temperature, because all sites are in close proximity (figure 1b). Thus, it seems likely here that temperature itself led to the evolution of countergradient variation in growth. It may be that this reduction in growth arises as a correlated response to selection on metabolic traits, but past work in other systems shows mixed evidence for temperature-induced countergradient variation in metabolism (e.g. [62,63]). Alternatively, warming-induced evolution of reduced growth early in life may result from a trade-off between growth rate and resistance to oxidative stress [64]. The elevated metabolism experienced by animals under warming may increase their exposure to harmful oxygen species [65], causing a shift along this trade-off towards slower growth. Regardless of the reasons for this reduction in growth, it is clear from our results that reduced individual growth rates can evolve quickly in response to temperature. This reduction in early growth, coupled with a likely decrease in somatic growth after maturity, is likely to contribute to reduced body size distributions at warmer temperatures in this species. Because countergradient variation in growth is common across a variety of taxa, and widely observed in fishes, these altered growth and development schedules may be important for predicting fisheries sustainability and yields in a warming world. Our data emphasize a role for both plasticity and evolution in contributing to reduced body size under warming. Field surveys indicated a strong reduction in offspring size at higher temperatures, that common rearing suggested was due to plasticity. Laboratory rearing in the sister species G. holbrooki also shows that higher rearing temperatures cause a plastic decline in size at maturity [66]. Therefore, mosquitofish, like many taxa, support the temperature–size rule of reduced stage-specific body sizes at higher rearing temperatures due to plasticity. However, our data also support a role for rapid evolution in contributing to these plastic size declines. First, evolution caused a reduction of juvenile growth rates. Second, evolution caused an increase in reproductive effort at small sizes. This increase in reproductive effort is likely to exacerbate the reduction in somatic growth rates after maturity. Altogether, this plasticity and evolution, coupled with the demographic effects of likely increases in mortality rates under warming, suggest that multiple processes may combine to produce a substantial decline in population body size distributions under warming [67]. However, to extend the trait evolution found here to population-level outcomes, future work should aim to (1) understand the combined effects of evolution and plasticity, including potential transgenerational plasticity not accounted for here [68], and (2) evaluate the role of warming for the evolution of body size, growth, and reproductive success in males as well. Overall, this study provides evidence that evolutionary adaptation to temperature itself is likely to contribute to reduced body size and growth rates over short time scales, compounding size reductions caused by other mechanisms. Body size is a key functional trait [1–6], and for fishes and other taxa, it is of key economic importance [69]. Although it is yet unknown whether the degree of evolution here is sufficient to significantly alter ecological dynamics, it is clear that this evolution can happen over the short time scales required to potentially affect these outcomes in the near-term future. Indeed, a young but growing literature suggests that rapid evolution and body size changes may significantly mediate the ecological consequences of warming [15,16,18,70–72]. Thus, if we aim to forecast the ecological consequences of warming for populations, ecosystems and society, we may need to incorporate body size and growth evolution into these models. Collections were approved by our institutional animal ethics committee (UCSC protocols PALKE-1311 and PALKE-1801) and the local wildlife agency (CADFW permit SC-12752). Data are available on the Dryad Digital Repository: https://doi.org/10.5061/dryad.cc2fqz63k [60]. D.C.F. led the study design, with contributions from E.R.M., M.T.K., K.S.S. and E.P.P. D.C.F. led collection of field data and wild fish, and performed common rearing (with help from F.J.A. and J.N.B.). D.C.F. developed the dissection protocol. A.N.H. led wild-caught trait measurements. D.A.A. led F1 trait measurements. D.C.F. and F.J.A. led F2 trait measurements. D.C.F. performed the statistical analyses. D.C.F. wrote the first draft of the manuscript, which was edited by all authors. We declare we have no competing interests. This research was supported by our affiliated institutions, the US National Science Foundation (a GRF to D.C.F., DEB 1457333 to E.P.P. and DEB 1457112 to M.T.K.), and the Royal Society of New Zealand Marsden Fund (16-UOA-023) to K.S.S., E.P.P. and M.T.K. E.P.P. received partial support from the Cooperative Institute for Marine Ecosystems and Climate. We thank S. Parmenter, P. Pister and C. Wickham for introducing us to these geothermal springs populations. For technical assistance, we thank T. Apgar, R. Franks and J. Velzy. For help with field collections, we thank K. Morrison, J. Penfield, S. Munch and B. Wasserman. For help with dissections and trait measurements, we thank E. Portillo, T. Amarnath, S. Bell, R. Robinson, E. Glensky, M. Haptonstall and M. De Aquino. For help with fish care during common rearing, we thank D. Ruiz, D. Weiler and T. Vance. Footnotes†Co-senior authorship. Electronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.4977722. © 2020 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original author and source are credited. References
Page 25Around 15% of human mortality around the world is caused by infectious and parasitic diseases [1]. This is relatively good news, as infectious diseases have probably been one of the greatest selection pressures on our species' evolution [2]. Given such pressures, one might expect evolution to have selected for abilities to detect and avoid infection threats. Indeed, existing work suggests people can use certain sensory information to accurately identify whether another person is infected. However, among the five major senses, hearing has been relatively unexamined in this context. Can people identify infectious disease from sounds such as coughs and sneezes? In this paper, we report four studies suggesting the answer to this question might, surprisingly, be no. Why might infectious disease detection abilities exist? The physiological immune system is an evolved mechanism for coping with pathogen threats. However, immune defences are energetically expensive, and they risk collateral damage. Given these issues, defensive psychological mechanisms that help avoid infection through behaviour would complement the physiological immune system. Such mechanisms have been referred to as the behavioural immune system [3,4]. Multiple lines of evidence indicate that organisms use sensory cues, specifically those tied to objects and behaviours historically associated with pathogen transmission (e.g. rotting material, faeces), to improve pathogen threat detection accuracy. Many species of animals respond to olfactory, visual, tactile and gustatory cues connoting parasitic or toxic dangers with aversive and sanitation-promoting behaviours [5–8]. Humans also use a variety of sensory cues to detect pathogen presence, although research with humans is relatively recent and commonly restricted to Western undergraduates. Moreover, human evidence is limited by relatively few studies within any single modality. Reviewing this work, with regard to sight, people were repulsed by faecal-shaped objects, despite explicit knowledge that they are non-infectious [9]. More generally, stronger disgust reactions emerged when perceivers saw images of body fluid-like stimuli relative to comparable stimuli without such an appearance (e.g. lesions with pus versus a burn scar, [10]). Recent research also found perceivers could distinguish above chance the faces of sick and healthy individuals ([11]; but see [12]), though visual cues may not be especially helpful for predicting a person's susceptibility to future infection [13,14]. With regard to smell, participants judged the odours of healthy individuals injected with lipopolysaccharide (which activates immune function) as less healthy [15] and engaged in prophylactic behaviours when exposed to faecal-smelling chemicals [16]. Additionally, Russian participants rated the body odours of those infected with gonorrhoea as more unpleasant [17]. Finally, with touch, participants rated objects that were wet and similar to biological material (e.g. a dough mixture, as opposed to cotton rope) as more disgusting and likely to produce illness [18], whereas inducing disgust in individuals led to enhanced skin sensitivity, ostensibly facilitating avoidance of infectious stimuli [19]. Detection of pathogen threats is important for many species. However, very little work has examined the role of auditory cues in this process. The ability to identify infected individuals by sound seems useful in avoiding infection, especially as it would allow detection from safer distances. Consistent with the potential relevance of sound, people were more likely to wash their hands after hearing belching sounds [20]. In addition, hearing others cough or sneeze increased perceptions of multiple health threats [21]. Although these findings require replication, it seems that sounds can elicit pathogen-avoidance responses and behaviours. But can people accurately detect pathogens through sound? From an error management perspective [22], people may possess evolved biases that limit accuracy. First, uncertainty afflicts interpersonal pathogen detection, and, second, asymmetric costs in detection errors exist. Mistaking an infectious person as non-infectious and potentially exposing oneself to harmful pathogens is probably costlier than mistaking a non-infectious person as infectious. Thus, people may be biased to judge coughs and sneezes as originating from infected rather than non-infected people. Such a bias—regardless of its potential origins [23]—would reduce the costlier error (similar to a smoke detector's false-positive bias). Here, we report a pilot and three studies that: (i) test whether people can accurately detect pathogen threats from cough and sneeze sounds, and (ii) test potential explanations for any detection abilities. Regarding detection abilities, perceivers may have lay theories about which sound dimensions cue pathogen threat, and these theories may help them reliably diagnose infection. For example, if disgust evolved to promote disease avoidance [10,24], and people are disgusted by sensory cues that indicate the presence of pathogens, then people may believe that disgusting features of cough and sneeze sounds can appropriately diagnose infection. We tested this both by directing participants to use their disgust reactions to make judgements about sounds that were infectious or non-infectious in origin and by examining associations between disgust ratings and accuracy. Finally, we also tested whether trait-level concerns about infection are associated with detection accuracy. An initial pilot study (see [25]) with 116 participants used similar methods to those described in study 1 below. After listening to and attempting to identify whether coughs and sneezes were infectious or non-infectious in origin, participants accurately identified 44% of the sounds' origins, 95% confidence interval (CI) [34%, 54%], a result not statistically different from chance (50%). In our primary studies, we expanded on this finding using a larger stimulus set and more varied measures. For studies 2 and 3, we preregistered research questions, predictions, sampling plans, exclusion criteria, and analyses on AsPredicted.org or the Open Science Framework. For these and additional analyses, see ([25] https://osf.io/4c7vr/). Previewing our findings, we find no evidence that perceivers can accurately detect (above chance) pathogen threat from sounds of coughs and sneezes, even though perceivers are highly certain in their judgements. We do find that the more disgusting that people perceive such sounds, the more likely they are to judge them as originating from infected others. However, attending to this disgust does not appear to improve pathogen threat detection accuracy. Using TurkPrime [26], we recruited 165 United States (US) participants from Amazon's Mechanical Turk system. Participants had completed at least 100 prior MTurk assignments with a 95% approval rate and were paid $0.90. Our final sample comprised 148 participants (mean age (Mage) = 37.02, s.d.age = 10.73, 83 women), which afforded 80% power to detect about an 11% difference from 50% (Cohen's h = 0.23; [27]). See [25] for detailed, preregistered exclusions based on hearing issues, medical training, and survey completion (studies 1–3). Conclusions in all studies are robust to these exclusions. Sound stimuli featuring coughs and sneezes were extracted from online, US-based videos (e.g. YouTube) (see [25]). We included different types of sounds to improve ecological validity, though we had no predictions about sound type differences. Targets who generated the infectious sounds self-reported with certainty experiencing sickness with an infectious disease (e.g. cold, flu). Targets who generated the non-infectious sounds responded to benign irritants (e.g. allergies, consumption of powdery spices, cotton swabs). We trimmed videos to 1–2 s audio clips featuring only the target sound. The full stimulus set comprised 20 coughs and 20 sneezes, with half of each sound type being infectious or non-infectious in origin. After responding positively to a volume check, participants were given instructions about sound identification. Infectious illness was defined as ‘an illness that can spread between people' and non-infectious was not defined. Sounds were presented in random order. After each clip, participants answered: (i) ‘do you think the sound is from a person with an infectious illness or a person with a non-infectious condition?' (infectious/non-infectious); (ii) ‘how certain are you that your above answer is correct?' (1/9 = not at all certain/very certain); and (iii) ‘how clear is the sound in this audio clip (how well could you hear it)?' (1/9 = not at all clear/very clear). Participants next completed a trait-level index of disease concern, the perceived vulnerability to disease questionnaire [28]. To save space, we report models using this scale in [25]. Finally, participants completed demographic items. In all studies, contingent on our dependent variable (e.g. 0/1 = incorrect/correct identification, 1/9 = not at all certain/very certain), we used the lme4 package in R [29] to fit either logistic or linear mixed effects models (deviating from our preregistered repeated measures ANOVA analyses). These models account for variability owing to sound origin (our key variable) while also accounting for sampling variability in participants and sound stimuli. However, the participant factor never accounted for meaningful variance (responses differed more within than between participants), so we excluded it in all reported models (see [25] for all models). We always included sound origin (non-infectious origin = −0.5, infectious origin = 0.5) as a fixed effect. Thus, unless otherwise noted, for all models using accuracy as the dependent measure (reported as percentage correct and log-odds), we included only stimuli intercepts. Importantly, because our diagnostic test was binary, overall accuracy is identical to the area under the receiver operating characteristic curve. By contrast, when we modelled Likert-like dependent measures (e.g. 1/9 = not at all certain/very certain), we specified random intercepts for participants and sound clips, and we specified random participant slopes for sound origin because we observed participants responding uniquely across the two sound origin categories. We describe additional fixed effect specifications in our Results sections. Because we made no predictions for sound type, we did not include it as a factor in our main analyses (see [25] for models including this factor). To increase confidence in our conclusions, we repeated the detection analyses using the subset of stimuli containing an explicit infection diagnosis. When comparing judgement accuracy for this subset to every possible non-infectious stimulus set of equal size, statistical conclusions did not differ from those involving the full stimulus set [25]. Thus, across studies, we report analyses using all stimuli. In addition, we had five research assistants code age, gender and race/ethnicity of the target individuals featured in the sound stimuli to evaluate as potential confounds. Including these variables in our regressions did not meaningfully affect coefficients nor conclusions, so they are not included in the analyses. Details are provided in [25]. Are people able to accurately identify infectious and non-infectious coughs and sneezes? We find no sufficient evidence that people can. On average, participants correctly identified 45% of the sounds, consistent with chance (50%), b = −0.22, 95% CI [−0.57, 0.13].1 The upper bound of our CI for overall accuracy translates to 53%, so we can reject overall accuracy above 53%. Accuracy also did not significantly depend on whether the sounds were truly infectious or non-infectious in origin, 95% CI [−0.69, 0.71] (figure 1). Similar to our pilot study, this low overall accuracy stems from the fact that participants more often classified non-infectious sounds as infectious (55% false positive rate/45% specificity) than infectious sounds as infectious (45% true positive rate/sensitivity). Figure 1. Plots visualize judgement accuracy by sound origin across studies (study 3 depicts separate panels by sound rating condition). The dashed lines represent chance levels, and error bars represent profile 95% confidence intervals based on the standard error of the sound origin difference. Points represent average accuracy for stimuli, ‘jittered' with random noise to make visible unique accuracy scores. (Online version in colour.)
Despite their accuracy being consistent with chance, perceivers were reasonably certain about their judgements, M = 6.68, 95% CI [6.48, 6.89] (above the mid-point of 5). In fact, subjective certainty negatively correlated with accuracy such that for every extra unit more certain participants were about their judgements, their odds of having correctly identified a sound's origin decreased by 5% (odds ratio (OR) = 0.95), b = −0.05, 95% CI [−0.08, −0.01], z = 2.71, p = 0.007. However, this association strongly depended on sound origin, b = 0.12, 95% CI [0.05, 0.18], z = 3.45, p < 0.001. Participants who reported higher certainty in judgements of sounds with infectious origins were not significantly more accurate than those who were less certain (z = 0.51, p = 0.607), but participants who reported higher certainty in their judgements about sounds with non-infectious origins were significantly less accurate than those who were less certain (z = −4.43, p < 0.001). This might indicate that perceivers are using non-diagnostic auditory cues to infer pathogen threat, a point we return to in the next study. The results of two initial studies suggest that perceivers may not be able to accurately detect pathogen threats from cough and sneeze sounds. One possibility is that individuals may be relying on misleading auditory information. To test this, in study 2, we focused on the role of disgust perceptions by instructing participants to rate how disgusting the sound stimuli were before identifying their origin. Given previous work suggesting that pathogen threat cues elicit disgust [10,18], attention to disgust might enhance accuracy in this context. A second issue is that perceivers may have lay beliefs about the natural frequency of infectious sounds in everyday life. In the current paradigm, half of the sound stimuli are infectious in origin, and half are non-infectious (though participants in study 1 were not told this). If people typically encounter different frequency distributions in their everyday lives, their judgements may reflect those more naturalistic distributions. Therefore, in study 2, we provided base rates by telling participants that half of the sounds were infectious in origin. One hundred and fifty people recruited from MTurk using the same methods in study 1 participated for $0.90. Our final sample comprised 146 participants (Mage = 35.14, s.d.age = 9.73, 57 women), which afforded us 80% power to detect about a 11% difference from 50% (Cohen's h = 0.23). We used the same stimuli and procedure from study 1, except for two key differences. First, before hearing the sound clips, participants read: ‘in total, you will listen to 10 infectious coughs, 10 non-infectious coughs, 10 infectious sneezes, and 10 non-infectious sneezes. These will be randomly ordered, so you will have to judge the type of each sound.' Second, before identifying each sound clip as infectious or non-infectious in origin, participants were instructed to rate ‘how disgusting do you find this sound?' (1/9 = not at all disgusting/very disgusting). Given perceiver knowledge of sound origin base rates and prior ratings of disgust, were participants able to accurately identify infectious and non-infectious cough and sneeze sounds? Still, we found no sufficient evidence for this ability. On average, participants correctly identified 42% of the sounds as either infectious or not, again not significantly different from chance2, b = −0.28, 95% CI [−0.61, 0.05]. The upper bound of the CI for overall accuracy translates to 51%, so we can reject overall accuracy above 51%. Also, like in study 1, accuracy did not significantly depend on whether the sounds were truly infectious or non-infectious in origin, 95% CI [−0.68, 0.64], p = 0.954 (figure 1). Similar to study 1, low accuracy stems from the fact that participants more often classified non-infectious sounds as infectious (56% false positive rate/44% specificity) than infectious sounds as infectious (44% true positive rate/sensitivity). Were disgust ratings associated with judgement accuracy? We find no sufficient evidence for an average association between disgust ratings and accuracy, 95% CI [−0.01, 0.05]. However, this association strongly depended on sound origin, b = 0.75, 95% CI [0.69, 0.81], z = 25.11, p < 0.001. For sounds with infectious origins, the more disgusting that participants perceived those sounds, the more accurately they identified them (i.e. judging them correctly as having an infectious origin). For sounds with non-infectious origins, the more disgusting that participants perceived those sounds, the less accurately they identified them (i.e. judging them incorrectly as having an infectious origin). These patterns are consistent with the hypothesis that disgust response is used as an index for pathogen presence: if they perceive a person's cough or sneeze as disgusting, they are more likely to judge it as having an infectious origin. As in study 1, participants were reasonably certain about their judgements, M = 6.08, 95% CI [5.80, 6.35] (above the mid-point of 5). For every additional unit of certainty, the odds of their accurately judging the sound's origin decreased by 4% (OR = 0.96), b = −0.04, 95% CI [−0.07, −0.01], z = 2.89, p = 0.004. Unlike in study 1, we found no sufficient evidence this association depended on sound origin, 95% CI [−0.03, 0.08]. Comparing study 2 to study 1, it seems that focusing on how disgusting cough and sneeze sounds are perceived does not significantly improve judgement accuracy for those sounds' origin. Perceivers did rate sounds they thought were disgusting as more likely to be infectious in origin. This may reflect perceivers' lay theories that more disgusting sounds are more likely to be infectious. However, study 2 lacked a control condition, which would allow us to directly test whether orienting participants to how disgusting the sounds were—compared to another subjective quality—improves accuracy. In study 3, we randomly assigned participants to rate sounds on either disgust or clarity before making origin judgements. Two hundred and twenty-four people recruited from MTurk using the methods from prior studies participated for $1.00. Our final sample comprised 211 participants (Mage = 37.01, s.d.age = 12.24, 108 women), which afforded us 80% power to detect about a 10% difference from 50% (Cohen's h = 0.19). Study 3 followed the same procedures as study 2 except we randomly assigned participants to rate disgust (n = 106) or clarity (n = 105) for each sound before judging its origins. Participants read: ‘research has shown that people are better at identifying the origin of sounds like the ones you'll hear when they pay attention to the sounds' (auditory clarity/disgusting quality). Therefore, when you attempt to identify each sound, pay special attention to (how clear the sound is/how disgusting the sound is) before making your decision.' For each sound clip, participants responded to the question ‘how [clear/disgusting] do you find this sound?' (1/9 = not at all [clear/disgusting]/very [clear/disgusting]). Are people able to accurately identify the origins of infectious and non-infectious cough and sneeze sounds when explicitly attending to how clear or disgusting the sounds are? On average, participants accurately identified sound origin 43% of the time, not significantly different from chance, b = −0.30, 95% CI [−0.66, 0.07]. The upper bound of the CI for overall accuracy translates to 52%, so we can reject overall accuracy above 52%. Also, like in studies 1–2, accuracy did not significantly depend on whether the sounds were truly infectious or non-infectious in origin, 95% CI [−0.67, 0.79] (figure 1). Similar to studies 1 and 2, low overall accuracy stems from participants more often classifying non-infectious sounds as infectious (57% and 56% false positive rates/43% and 44% specificity for disgust and clarity conditions) than infectious sounds as infectious (45% and 43% true positive rates/sensitivities for disgust and clarity conditions). Were clarity and disgust ratings associated with accuracy? We found no significant average association, but the association between sound rating and accuracy significantly depended on both sound origin and rating condition (a three-way interaction) (figure 2), b = 1.02, 95% CI [0.91, 1.14], z = 17.88, p < 0.001. In the disgust rating condition, the more disgusting participants perceived sounds with infectious origins, the more accurately they identified them (z = 17.98, p < 0.001), but the more disgusting participants perceived sounds with non-infectious origins, the less accurately they identified them (z = −18.14, p < 0.001). This is similar to the pattern found in study 2. By contrast, we found no sufficient evidence for associations between perceived clarity and accuracy (zs = 0.07 and −0.25 and ps = 0.947 and 0.800). Figure 2. Plots visualize the relationship between sound rating (clarity or disgust) and judgement accuracy by sound origin (infectious or non-infectious targets). Higher ratings of disgust correlated positively with accuracy for infectious sounds but negatively with accuracy for non-infectious sounds (clarity ratings had no significant effect). Dashed lines represent adjusted average accuracy for each condition. Small points represent raw judgements ‘jittered’ with random noise to make visible unique judgements. Ribbons around the lines represent 95% confidence intervals based on the three-way interaction (see effects package in R [30]). (Online version in colour.)
As in studies 1–2, participants were reasonably certain about their judgements, M = 6.35, 95% CI [6.14, 6.56] (above the mid-point of 5). For every additional unit more certain participants rated their judgement, the odds of their accurately judging sound origin decreased by 4% (OR = 0.96), b = −0.04, 95% CI [−0.065, −0.01], z = 2.97, p = 0.003. Can perceivers detect pathogen threats from cough and sneeze sounds? Given our hypothesis that this capacity would be adaptive for limiting pathogenic exposure, and existing work indicating similar capacities with other sensory modalities, we predicted that people would be able to accurately detect pathogen threat using auditory cues. Across four studies, we found no support for this prediction. Moreover, there was no evidence that accuracy improved when participants knew the true number of infectious sounds in advance or when participants focused on how clear or disgusting they perceived the sounds. Despite this poor overall accuracy, perceivers consistently reported reasonable certainty in their judgements. Poor accuracy notwithstanding, we identified one subjective dimension used to identify the origin of sounds. In studies 2 and 3, the more disgusting that participants rated sounds, the more likely they were to judge a sound as infectious in origin, regardless of whether it truly was. This was not the case for another subjective dimension—sound clarity. It seems that perceivers possess the lay theory that what disgusts them is likely to represent a disease threat, potentially leading them to exhibit biases to avoid interactions with others who make disgusting but non-infectious noises. Given these findings, should we conclude that people possess no auditory disease detection mechanism? Perhaps coughs and sneezes are such strong and consistent physiological reflexes that any variety of causes, infectious or not, will produce very similar sounds. If so, the necessary sound variation would not have been available to natural selection processes (or perceivers). Such sound variation could be very limited, but recent work suggests that this variation is available (e.g. to statistical learning algorithms; [31]). However, human hearing mechanisms may not be able to use it reliably, even with clinical training (e.g. [32]). Another possibility is that people integrate available sickness sound information with other sensory information (e.g. [33]) rather than perceiving it in isolation. Accuracy may improve when people hear sounds alongside other sensory cues (e.g. seeing someone sneeze, hearing someone talk with a hoarse voice). Of course, the same is true for other sensory modalities, including ones that have been tested in isolation and found to predict accurate identification (e.g. [11]). A final possibility is that our stimuli may not represent the relevant range of sounds that perceivers encounter in natural settings. We tried to address this by using a range of sound types (i.e. coughs and sneezes), but our set may have nonetheless been limited in type, quality, and breadth of eliciting conditions. People certainly attend and react to auditory information in contexts related to infectious disease threats [20,21]. However, our data suggest they are poorly able to distinguish infectious cough and sneeze sounds from benign ones. From an error management perspective, biases to presume that coughs and sneezes indicate pathogen presence could be functional if sound origin is uncertain and the costs of mistaking benign coughs and sneezes for infectious ones are lower than the costs of mistaking infectious ones as benign. However, this also does not appear to fully explain our results. Individual differences in self-perceptions of vulnerability to disease were not significantly associated in a consistent manner with inferences of infectious threat (see [25]), as this perspective might predict. Work on the psychological mechanisms that manage infectious disease threat has recently flourished, coupled with advances in our understanding of the selection pressures that pathogens have exerted on human evolution. Despite this growing literature, relatively little research has focused on the manner by which people accurately identify pathogen threats. The work presented here on auditory detection expands on this emerging literature. Relevant information about sickness is carried through sound, and perceivers believe this information to be both useful and relatively safe [34], making it useful to know whether auditory information is interpreted accurately. Adaptive explanations also exist to predict either accurate or biased perception of infection threats through sounds. Though the nascent state of the literature makes this investigation largely exploratory, the current work speaks to these alternative hypotheses and advances future theorizing about pathogen threat processing. In conclusion, we find no evidence that perceivers can reliably detect pathogen threats from cough and sneeze sounds, even though they are reasonably certain they can. Perceivers seem to use the disgustingness of these sounds to infer pathogen threat but paying attention to their disgust does not significantly improve accuracy. Thus, the next time you hear someone cough or sneeze, perhaps leave the diagnosis to the doctor. –IRB Exemption Status– The University of Michigan IRB for Health Sciences and Behavioral Sciences has reviewed and approved exempt status (HUM00125412) to the studies presented in the manuscript. At the beginning of each study, participants gave informed consent. We include all preregistrations, study materials, de-identified data, analysis code and supplementary write-ups in our Open Science Framework (OSF) Project: https://osf.io/4c7vr/. J.A. generated the research question, collected the stimulus sets and conducted the pilot study and study 1. All authors contributed in generating research questions, study designs and data collection for studies 2–3. N.M.M., O.S. and J.A. contributed in preparing and analysing data and interpreting results. O.S., J.A., N.M.M. and I.M.W. drafted the manuscript's Introduction and General Discussion. N.M.M., O.S., J.A. and I.M.W. drafted the manuscript's Methods and Results. N.M.M. prepared data, analyses code and supplemental repository. All authors gave final approval for publication and agree to be held accountable for the work performed therein. We declare we have no competing interests. We received no funding for this study. Footnotes1 We computed 95% confidence intervals using the likelihood profile method [29]. 2 Here and across studies, overall accuracy falls below 50%. Adjusting for sound type (cough or sneeze), target demographics, and sound disgustingness did not substantively change this pattern. References
Page 26Allotherians are an extinct group of mammaliaforms, primarily known from the Mesozoic, that are currently the subject of conflicting phylogenetic hypotheses (figure 1). Allotherians share a number of dental apomorphies, most notably postcanines with multiple cusps in longitudinal rows (superficially resembling those of some therian mammals, such as rodents), and they include haramiyidans, multituberculates and gondwanatherians [1–6]. Some phylogenetic analyses have supported monophyly of Allotheria, within (crown-clade) Mammalia [2,7–10] (figure 1, topology 1). Conversely, others have recovered haramiyidans outside Mammalia, but with multituberculates remaining within Mammalia [3,11,12] (figure 1, topology 2a), suggesting that allotherian dental apomorphies have evolved more than once. Finally, two studies recovered diphyletic haramiyidans, with the euharamiyidans forming a clade with multituberculates within crown mammals, but the Triassic species Haramiyavia and Thomasia falling outside the crown group [13,14] (figure 1, topology 2b). Figure 1. Summary of the main hypotheses for the relationships of ‘allotherians’. Numbers refer to the number of different independent clades of ‘allotherians’. Arrows indicate the mammalian crown node. Three independent clades is a novel result from this study. (Online version in colour.)
Monophyly versus polyphyly of Allotheria has major implications for our understanding of Mesozoic mammal evolution, leading to different scenarios for the evolution of numerous dental and skeletal features, including the so-called Definitive Mammalian Middle Ear, in which the angular, articular, prearticular, and quadrate have become entirely auditory in function, and are fully separated from the jaw joint [9,15,16]. It also affects interpretations of the age of Mammalia: if Late Triassic haramiyidans such as Haramiyavia and Thomasia fall within the crown-clade, then the split between monotremes and therians must be at least this old [8]; if they fall outside the crown-clade, this split could be considerably younger, as it would render Asfaltomylos and Henosferus (which appear to be early relatives of monotremes) from the Middle Jurassic of Patagonia the oldest known members of the crown-clade [17]. Another fossil mammal that has been the subject of recent discussion is the eutherian Juramaia sinensis from the Middle–Late Jurassic (164–159 Ma) Linglongta Biota (the younger of the two phases composing the Yanliao/Daohugou Biota) from the Lanqi/Tiaojishan Formation of China [18,19]. Based on its known morphology, Juramaia has been argued by some authors [20–22] to be ‘unexpectedly advanced’ for its age, as it closely resembles eutherians from the much younger (ca 126 Ma) Jehol Biota [20,22]. By contrast, the same has not been argued for other mammaliaforms from the Yanliao Biota. However, whether or not the known morphology of Juramaia is ‘unexpected’ given its age has never, to our knowledge, been quantitatively tested. Tip-dated phylogenetic methods [23], which include morphological and stratigraphic data in a single analytical framework, are a promising avenue to investigate these issues. The wide time difference between the earliest known haramiyidans (Late Triassic) and the oldest known multituberculates (Middle Jurassic) [5,24] suggests that their similarities may be the result of convergent evolution, and incorporating stratigraphic data into phylogenetic analysis means that this temporal disparity is taken into account [25]. Another use of tip dating is to use the morphological data to inform the ages of fossils with uncertain dates [26,27]. Given that the known morphology of Juramaia has been identified as ‘unexpectedly advanced’ [20–22], it can be used to test whether tip dating continues to support a Middle–Late Jurassic age when its age is allowed to vary. Here, we apply tip dating to recent datasets of Mesozoic mammals to investigate the relationships of the haramiyidans, and to test the congruence between the known morphology and age of Juramaia and other Yanliao mammaliaforms. Our focal dataset was taken from Huttenlocker et al. [3], which comprises 538 morphological characters scored for 125 mammaliaforms and non-mammaliaform cynodonts. Because the sampling of Cenozoic taxa in this dataset was extremely sparse relative to Mesozoic taxa, extant and Cenozoic fossil taxa were pruned from the dataset, and invariant characters in this reduced dataset were deleted, leaving 96 taxa and 507 characters. Tip-dated Bayesian analyses were performed in BEAST v. 2.5.2 [28]. The Markov model for variable characters (hereafter Mkv) was used [29], with a gamma distribution (with four rate categories) to account for rate variation across sites. Characters were partitioned according to the number of character states. The clock model was an uncorrelated lognormal clock [30], and the tree prior was a sampled-ancestor fossilised birth–death model [31]. Tip dates were assigned uniform priors across the range of uncertainty for each taxon. The analysis was run for 1 billion generations, sampling every 500 000. Convergence of four independent runs was confirmed in Tracer [32], and the R package RWTY [33]. To investigate conflicts between the different parts of the dataset of Huttenlocker et al. [3], and further test allotherian relationships, the following character subsets were analysed individually: craniodental, dental only, and postcranial only. Undated Bayesian analyses were performed in MrBayes [34], again using the Mkv model with a gamma distribution (with four rate categories) to account for rate variation across sites. Four independent runs, each with four chains, were run for 10 million generations, sampling every 5000. Parsimony analyses in TNT [35] employed new technology search, using sectorial search and tree fusing with default settings for 1000 random addition sequences, followed by tree bisection and reconnection swapping to fully explore tree islands. We also ran a constrained parsimony analysis with a negative constraint on haramiyidan monophyly. To further test the extent to which tip dating could overturn topologies supported under other methods, similar tip-dated analyses were run on the datasets of Krause et al. [2] and Wang et al. [15], both of which originally recovered a monophyletic Allotheria. Extant taxa were pruned, as above, resulting in datasets of 81 taxa, 448 characters and 89 taxa, 473 characters, respectively. Tip-dated analysis of the Krause et al. [2] dataset showed very poor mixing (caused by alternative likelihood peaks representing monophyly or polyphyly of Allotheria) and was therefore run for 32 independent runs, each of a billion generations, to obtain reliable estimates of the relative posterior probabilities of the two phylogenetic hypotheses. Results from each run were thinned (sampling every 5 million generations) and, following removal of a 50% burn-in from each run, combined for further analysis. We also ran an analysis of the Huttenlocker et al. [3] dataset with a wider prior age range for Juramaia. This represents a quantitative test of the ability of tip dating to infer the age of Juramaia based on its known morphology. The tip age prior for Juramaia was modified to a Laplace distribution centred on 161 Ma, with a scale parameter of 8. This represents a strong prior expectation that Juramaia is Jurassic in age (with 90% of the prior probability density between 143 and 179 Ma), but owing to the absence of hard maximum or minimum bounds, dates outside this range are permitted. The other taxa from the Yanliao Biota in this dataset (Agilodocodon, Arboroharamiya, Castorocauda, Docofossor, Maiopatagium, Megaconus, Pseudotribos, Rugosodon, Shenshou, Vilevolodon, Xianshou linglong and Xianshou songae), were given the same Laplace distribution prior in separate analyses, to test the effectiveness of this method. Extraction of branch rates from the consensus trees for plotting (electronic supplementary material, figure S13) used the R package OutbreakTools [36]. Tip-dated analysis of our focal dataset, modified from Huttenlocker et al. [3], resulted in allotherian taxa falling into three separate clades (figure 2). The Late Triassic haramiyidans Haramiyavia and Thomasia are placed outside Mammaliaformes, in a strongly supported clade with tritylodontids (posterior probability (PP) = 0.91). The Middle Jurassic euharamiyidans, Early Cretaceous hahnodontids, and the Late Cretaceous Madagascan gondwanatherian Vintana, by contrast, collectively form a strongly supported clade (PP = 1.00) within Mammaliaformes, although our phylogeny is insufficiently well resolved to indicate whether or not this is within crown-clade Mammalia. Finally, the multituberculates form a third strongly supported clade (PP = 1.00), within Mammalia. Figure 2. Fifty per cent majority rule consensus tree for Mesozoic mammals from a tip-dated analysis of the dataset in Huttenlocker et al. [3]. Allotherians (Haramiyavia, Thomasia, euharamiyidans, multituberculates, hahnodontids and the gondwanatherian Vintana) are in green. Branch widths are proportional to posterior probability (between 0.5 and 1.0). Labelled branches (1–3) indicate the branches used for the rate analysis in the electronic supplementary material, figure S13. (Online version in colour.)
Both undated Bayesian and parsimony analysis recovered monophyletic Haramiyida (table 1). Parsimony analysis with a negative constraint on haramiyidan monophyly (i.e. preventing Haramiyavia and Thomasia from forming a clade with euharamiyidans) produce trees that are only two steps longer (representing just a 0.1% increase in tree length) than the unconstrained trees. Constrained and unconstrained trees were not significantly different (p = 0.87) under the Templeton test [37].
Support for monophyly of Allotheria and of Haramyida is driven by dental characters, and it should be noted that Thomasia is known only from isolated teeth and that Haramiyavia is also represented almost exclusively by dental characters. Analysis of craniodental or dental only character subsets led to allotherians falling into progressively fewer separate clades across tip-dated, undated and parsimony methods (table 1; electronic supplementary material, figures S3–S5). Strong support for allotherian polyphyly (i.e. three independent clades) is only found under tip dating on the full dataset, whereas all methods support allotherian monophyly when dental characters are considered in isolation. Tip-dated analysis of postcranial characters only also recovers separate euharamyidan and multituberculate clades (electronic supplementary material, figure S5), but Haramiyavia, Thomasia, hahnodontids and Vintana could not be included in this analysis as postcranial remains have not been described for them [13,38]. Tip dating using the Wang et al. [15] dataset recovered a diphyletic Haramiyida (electronic supplementary material, figure S6), with euharamiyids and multituberculates forming a clade distant from Haramiyavia + Thomasia (figure 1, topology 2b). The dataset of Krause et al. [2] led to a more complex result, as the sample of post-burn-in trees includes some topologies in which Allotheria is polyphyletic and others in which it is monophyletic. This analysis showed ‘twin peak’ behaviour of the prior and likelihood traces (electronic supplementary material, figure S7). These peaks correspond to the two different tree topologies regarding Allotheria. One peak, where the Late Triassic Haramiyavia and Thomasia formed a clade with other allotherians (essentially the parsimony result) had a low prior (or tree model likelihood) but a high likelihood (electronic supplementary material, figure S8–S9). The other peak, which had Haramiyavia and Thomasia closer to the root of the tree, and separated from other allotherians, had a higher prior and lower likelihood (electronic supplementary material, figure S10). Overall, allotherian monophyly remained the preferred hypothesis, found in 73% of the posterior sample, compared with 27% showing polyphyly of Allotheria. Arboroharamiyavia, the only euharamiyidan included in the Krause et al. [2] dataset, was always recovered with multituberculates. A constrained parsimony search revealed that polyphyly of Allotheria requires four additional steps (a 2.23% increase in tree length) compared to the unconstrained analysis (which recovers allotherian monophyly). However, constrained and unconstrained trees were not significantly different (p = 0.68) under the Templeton test [37]. Rerunning the analysis on the Huttenlocker et al. [3] dataset without a hard upper or lower bound on the age of Juramaia had no effect on the recovered relationships of haramiyidans and multituberculates: haramiyidan diphyly (and allotherian triphyly) was still recovered (electronic supplementary material, figure S11). Strikingly, however, this analysis revealed a strong signal in the data supporting a post-Jurassic age for Juramaia (figure 3). The mean estimated age for Juramaia was 123.5 Ma, almost exactly the same as the age of the Jehol Biota, from where several fossil eutherians are known that are morphologically similar to Juramaia [22]. The 95% highest posterior density (HPD) interval was 106.3–137.6 Ma, entirely within the Early Cretaceous. This contrasts with the results from the other Yanliao Biota mammaliaforms. When these were assigned the same Laplace distribution age prior as Juramaia, the resulting age estimates were always Jurassic. Megaconus resulted in the most inaccurate age estimate (mean 173.6 Ma), but the 95% HPD interval (154.8–194.8 Ma) comfortably overlapped the true age of the Yanliao Biota. For all other taxa, mean age estimates were between 156.8 Ma (Rugosodon) and 164.1 Ma (Castorocauda) and 95% HPD intervals fell between 142.6 Ma (lower bound for Rugosodon) and 182.4 Ma (upper bound for Maiopatagium). The Juramaia result may be partly driven by low sampling of eutherians during the Early Cretaceous (electronic supplementary material, text; figure S12): estimating the age of Rugosodon after deleting the similarly aged multituberculate Kuehneodon and plagiaulacids resulted in a wide age estimate (95% HPD 114.4–164.5 Ma). Figure 3. The morphological data have a strong signal towards an Early Cretaceous rather than a Jurassic age for Juramaia. Age estimates for members of the Yanliao Biota, analysed with a Laplace distribution prior centred on 161 Ma, therefore representing a conservative test. Data for Juramaia strongly contradicts the prior, in contrast to the other Yanliao mammaliaforms, all of which retain a Jurassic age. (Online version in colour.)
The age of Juramaia also has a significant effect on estimated rates of evolution (electronic supplementary material, figure S13a). When Juramaia is assigned its correct, Middle–Late Jurassic age, rates of evolution on the branch leading to crown Theria, and the branch leading to Eutheria, are estimated to be the highest across the entire tree and nearly 10 times higher than the average for all branches, as previously reported by Close et al. [39]. The rate on the branch leading to Eutheria excluding Juramaia is however very low, suggesting a 50-fold decrease in evolutionary rates in eutherians across the Jurassic–Cretaceous boundary. However, when the age of Juramaia is allowed to vary (resulting in the estimation of an Early Cretaceous age), rates of evolution on these three branches are far more similar, resulting in approximately constant rates during early eutherian evolution (electronic supplementary material, figure S13b). The results of our tip-dated analysis of the Huttenlocker et al. [3] dataset suggest that the dental similarities proposed to unite Allotheria are homoplastic, and that they evolved at least three times independently: once in the common ancestor of Haramiyavia + Thomasia and tritylodontids, once in the common ancestor of euharamiyidans, hahnodontids and gondwanatherians, and once in multituberculates (contra [4,5,8,15,40]). Notably, a recent study found that dental characters in mammals are more prone to homoplasy than characters from the rest of the skeleton [41]. Our results are congruent with recently discovered morphological differences between Triassic haramiyidans and the euharamiyidans. In particular, Haramiyavia retains a prominent postdentary trough [13], a plesiomorphic feature indicating that it lacked fully detached ear ossicles, whereas in most euharamiyidans (with the notable exceptions of Megaconus and Vilevolodon [12,16,42]) this trough is either very small or absent [7–10,16]. In some ways, our results represent a compromise between differing views on whether haramiyidans are crown- or stem-mammals: euharamiyidans fall within or near the crown-clade, whereas Haramiyavia + Thomasia fall outside. Our analysis places Haramiyavia and Thomasia in a clade with tritylodontids, a result that may be the result of insufficient sampling of non-mammaliaform cynodont characters and taxa, and which we consider in need of further testing (see detailed discussion in the electronic supplementary material). The recovered phylogenetic relationships of allotherians depend on both the dataset and the method used. Tip-dated methods invariably push the results towards splitting up the allotherians, but the extent of this depends on the data matrix. For the Krause et al. [2] and Wang et al. [15] datasets, which originally recovered allotherian monophyly (figure 1, topology 1), tip dating leads to increased support for two independent lineages (figure 1, topology 2b), a topology possibly supported by recently discovered morphological similarities between early multituberculates and euharamiyids [24]. For the dataset from Huttenlocker et al. [3], which originally recovered separate haramiyidans and multituberculates (topology 2a), tip dating decisively supports three independent lineages (topology 3). The relative influence of stratigraphic and morphological data in tip-dated analyses remains an underexplored issue. Tip dating of the Huttenlocker et al. [3] dataset results in strong support for polyphyly of Allotheria, including diphyly of the haramiyidans, a result that requires only two additional steps under parsimony. By contrast, the dataset of Krause et al. [2] has stronger morphological support for allotherian monophyly. Analysis of this dataset flips between allotherian polyphyly and monophyly, and allotherian polyphyly requires four additional parsimony steps over monophyly. In the case of the Krause et al. [2] dataset, the stronger morphological signal for allotherian monophyly is therefore not fully overruled by the stratigraphic evidence. These results suggest that the stratigraphic data only become influential on tree topology when morphological support for conflicting topologies is weak. The effect of stratigraphic age on haramiyidan relationships is analysed quantitatively in the electronic supplementary material. For some datasets at least, Bayesian tip dating appears to perform relatively well at estimating the ages of tips when treated as unknown [26], although 95% HPDs can be wide [43]. However, in this case, this method failed to accurately identify Juramaia as Middle–Late Jurassic in age, confirming that this taxon is characterised by a morphology that is unusually derived given its age. The Jurassic age of Juramaia suggests unusually rapid rates of evolution at the base of therians and eutherians, followed by a 50-fold rate decrease and a period of exceptionally slow eutherian morphological evolution during the Early Cretaceous [39]. The Juramaia result requires further scrutiny owing to low sampling and phylogenetic uncertainty of early therian mammals (electronic supplementary material, text; figure S12). Our result is largely driven by two taxa, both of which are known from single specimens: Juramaia and Eomaia. The highly incomplete record of early eutherians [22] makes it difficult to reach robust conclusions regarding the macroevolution of the group, and these may change with future discoveries. Juramaia has also been considered to be a stem therian by some authors [44], a phylogenetic position that would be more consistent with its age. Finally, Sinodelphys has recently been proposed to be a eutherian rather than a metatherian [22]. If this is the case, it could alter branch length estimates, and influence inferred patterns of early eutherian evolution. Full data, analysis code, files and results are available as part of the electronic supplementary material. B.K. and R.M.D.B. designed the study and wrote the paper. B.K. performed analyses and produced figures. We declare we have no competing interests. We have received no funding for this article. B.K. thanks Martin Rücklin for generously supporting the completion of this paper. We thank Brian Davis, Guillermo Rougier, Erik Seiffert, Ian Corfe, Mike Lee, David Grossnickle, Lucas Weaver and an anonymous reviewer for their insightful and constructive comments that much improved the paper. TNT is made freely available through the Willi Hennig society. FootnotesElectronic supplementary material is available online at https://doi.org/10.6084/m9.figshare.c.4994606. © 2020 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original author and source are credited. References
|
What feature is unique to Chytrids compared to other fungi?
Related Posts
Copyright © 2024 SignalDuo Inc.
