Why is Madagascar special? The extraordinarily slow evolution of pelican spiders (Araneae, Archaeidae)

MORPHOLOGY AND DNA SEQUENCE COLLECTION

To examine archaeid relationships, we included 81 terminals representing 52 archaeid species from Madagascar, Australia, and southeastern Africa, comprising the four known extant genera: Eriauchenius (Cambridge 1881), Austrarchaea, Afrarchaea (Forster and Platnick 1984), and Zephyrarchaea (Rix and Harvey 2012a) as well as the monophyletic “Gracilicollis Group” (Wood 2008) that is currently considered part of Eriauchenius. Of the known Madagascan species, 18 of 21 described are included, plus an additional 15 undescribed, for a total of 33 species (out of 37 known). Twelve African (out of 12 known species) and seven Australian (out of 38 known species) are included. Although we include relatively few Australian species, we were able to include at least two species from each of the three major Australian clades (Rix and Harvey 2011, 2012a,b). Five fossil archaeid taxa, representing five of 11 fossil genera, were included using only morphological characters. Archaeids belong to the Palpimanoidea (Wood et al. 2012): the outgroup taxa include 13 terminals representing the remaining four Palpimanoidea families, the family Austrochilidae, and the tree is rooted with the Haplogynae Segestriidae, for a total of 99 terminals (Table 1).

The morphological characters from Wood et al. (2012) were used to score all species and outgroups. Molecular data for the austrochilid was acquired from GenBank, whereas the molecular data for the remaining specimens were gathered following the methods of Wood et al. (2012). A suite of primers was used to amplify a portion of the mitochondrial protein coding gene Cytochrome c Oxidase subunit 1 (COI), the nuclear protein-coding gene Histone-3 (H3), and the ribosomal nuclear genes 28S and 18S. Some specimens sequenced for this study were difficult to amplify and/or sequence for some markers, so that a few taxa are incomplete for some regions or markers (see Table 1). All DNA sequences have been deposited in GenBank under accession numbers KP201206-KP201459 (Table 1).

Nonprotein coding genes were aligned using the Mafft (Katoh et al. 2002) online server (http://mafft.cbrc.jp/alignment/server/) using the E-INS-i strategy, which operates best on sequences with conserved domains and long gaps. The gap open penalty and the offset value were set to their defaults of 1.5 and 0.123, respectively. The best-fit substitution model and the data partitions were estimated for the molecular data using PartitionFinder version 1.1.0 (Lanfear et al. 2012) for 28S, 18S, and each of the three codon positions in the protein coding genes, using greedy search for MrBayes models, using BIC criteria, and resulting in the final concatenated total evidence dataset having the following four partitions: morphology; COI codon position 3; COI codon positions 1 and 2 and H3; 18S and 28S. For the morphology partition, the standard discrete Markov (Mkv) model (Lewis 2001) was used with rates set to gamma. The best-fit models for the remaining three partitions were GTR + G, GTR + I + G, and K80 + I + G.

PHYLOGENETIC ANALYSIS AND DIVERGENCE DATING

Analyses were performed using Bayesian methods on the morphological matrix, the molecular concatenated data, and the total evidence data containing the molecular and morphological characters. Analyses were implemented in MrBayes version 3.2.2 (Ronquist and Huelsenbeck 2003) using four chains, with the analysis run twice simultaneously, and the starting trees randomly generated. Analyses were run for 10–20 million generations, depending on how quickly the analysis converged, with sampling every 1000th generation. All analyses were checked to ensure that the deviation of split frequencies was below 0.01. The two simultaneous analyses were evaluated for convergence using Tracer version 1.4 (Rambaut and Drummond 2007). The burn-in value was visualized and determined by summarizing posterior distributions of scalar values, resulting in a final consensus tree with node support expressed as posterior probabilities (pp).

To examine the timing of divergence, a time-calibrated phylogeny was created by treating archaeid fossil taxa as noncontemporaneous tips following the methods of Pyron (2011) and Wood et al. (2013). For each terminal fossil, the geological stage and reference are listed: (1) Archaea paradoxa, Baltarchaea conica, and Myrmecarchaea sp., from Baltic amber, Eocene, Lutetian, 44–49 Ma (Penney et al. 2011); (2) Burmesarchaea grimaldii, from Burmese amber, Cretaceous: Cenomanian-Turonian, 88–95 Ma (Penney 2003); (3) Patarchaea muralis, compression fossil, Middle Jurassic (Chen et al. 2004; Gao and Ren 2006), 161–176 Ma (based on www.geosociety.org/science/timescale/ last accessed August 8, 2014). The fossil age input was the median of the estimated range, for example, 46.5 Ma for the Baltic amber fossils. The root age was constrained, being treated as a normal distribution with a mean of 225 Ma with soft upper and lower bounds (standard deviation = 30; 5–95% bounds = 175.7–274.3 Ma), based on the age of the oldest Araneomorphae fossil (Selden et al. 1999), which is also close in age to the oldest Mygalomorphae fossil (Selden and Gall 1992; Penney et al. 2003). The breadth of this prior constraint was intentionally large to contain the true age of Araneomorphae divergence and so as not to bias the study.

We estimated the mean node ages and their 95% Bayesian credible interval (CI) using a relaxed clock model implemented in BEAST (Drummond and Rambaut 2007). The taxa, molecular markers, and partitions used for the phylogenetic analysis were also used for the BEAST analysis except duplicate species were pruned so that only one representative per species remained. The same partition models from the Bayesian phylogenetic analysis were used except that partition 2 and 3 (representing COI and H3) were changed to the simpler Hasegawa–Kishino–Yano (HKY) model. This was done to provide an analysis with a much lower chance of convergence problems and parameter nonidentifiability as these partitions were conserved, not having enough variation for their parameters to reach convergence. The molecular clock model was set to relaxed, uncorrelated lognormal and the tree prior was set to speciation, birth–death process. Five MCMC analyses were run for 30 million generations, sampling the chain every 1000 generations. Log files were visualized in Tracer version 1.4 to ensure that the effective sample size of the combined log files reached 200 for all parameters (Drummond et al. 2006). The burn-in was removed from each independent run and the resulting tree files were combined. Because mitochondrial DNA markers can lead to age overestimations under certain conditions (Near et al. 2012), this analysis was repeated excluding COI. Additionally, the original analysis was also performed using the independent gamma rates (IGR) model in MrBayes version 3.2.2 (Ronquist et al. 2012b) following the methods of Ronquist et al. (2012a), however the results were not realistic (see Supporting Information).

MEASUREMENTS AND HABITAT DATA

We measured seven morphological traits of the cephalothorax and one measurement of standard length per species (Fig. 2). The specific traits measured were: (1) carapace length; (2) carapace constriction, defined as the depth from anterior to posterior of the narrowest portion of the “neck” in the lateral habitus view; (3) carapace angle, defined as the angle between the posterior edge of the lateral side of the carapace (the portion above coxae II and III) and the anterior edge of the “neck” in the lateral view; (4) anterior “head” length, defined as the length between the clypeus and the posterior edge of the “head,” taken perpendicular to the cheliceral foramen in the dorsal view; (5) “neck” length, defined as the length from the bottom edge of the cheliceral foramen to the bottom edge of the carapace, taken in the anterior view; (6) chelicerae length; (7) length of femur I. Carapace width was used as the measurement for size. These traits were selected because of their ecological and functional significance in predatory attacks (Wood et al. 2007; Wood 2008). Measurements were taken, when possible, for five males and five females per species, and all traits were natural log transformed and averaged per species. The effect of size was removed by performing a phylogenetic linear regression on each measurement against the size measurement, and then retaining the residual values for each species (Revell 2009). The size-corrected residuals were used as inputs for a phylogenetically corrected principal components analysis and the resulting eigenstructure and scores were extracted for each species. Even though Australian species are undersampled, we feel that we have adequately sampled the morphospace, as we have a range of morphologies including long and short “necked” species from the three major Australian clades (Rix and Harvey 2011, 2012a,b).

Habitat data were used to examine whether Madagascar species have a greater rate of habitat evolution than their non-Madagascan relatives. Habitat was assessed based on the collection methods used to obtain the specimens, which explicitly records habitat of occurrence based on the techniques used. Habitat was broken into two categories, (1) up in the vegetation or (2) down on the forest floor, per specimen, and was then treated as a continuous trait, per species, of the percentage of time collected in the vegetation. A specimen was recorded as being in the vegetation if it was collected either by beating vegetation or by hand in the vegetation, and as being on the ground if it was collected from either pitfall traps, sifting litter, sweeping, or turning over rocks and logs. Per species, these values were turned into a percentage by dividing the specimens collected in vegetation by the total number of observations. Collection methods were examined mostly from the California Academy of Sciences (CAS) collection, but also from field expeditions made during the course of this study, and from the literature. There was an average of 19 observations per species with a range of 1–215. The species that did not have any recorded habitat data were pruned from the phylogeny for any subsequent analysis that examined habitat. Our characterization of habitat is a simplified version and thus, an underestimation of true habitat diversity: observations of HMW suggest there may be more habitat niches, however, currently, we do not have enough data to test this. Eighteen species in the dataset had less than five habitat observations. Although this lack of data is acknowledged, we still believe there are real differences between habitats among species.

LINEAGE AND TRAIT DIVERSIFICATION

Rate of lineage diversification was tested using the gamma statistic (Pybus and Harvey 2000) to examine how lineage diversification proceeded through time. Negative values of gamma indicate that internal nodes are disproportionally placed closer to the root than expected under the Yule model of speciation, suggesting that diversification rates are slowing over time. The Monte Carlo constant rates test (MCCR), which takes incomplete lineage sampling into account, was implemented in the R package LASER (Rabosky 2006) to identify the critical value for rejecting a constant diversification rate. The MCCR test was performed 5000 times for 37 total taxa with four taxa randomly removed from the phylogeny, mimicking the incomplete taxa sampling of the empirical Madagascar archaeid dataset (non-Madagascan archaeids were pruned from the phylogenies). Then the gamma statistic was calculated for 500 randomly sampled phylogenies from the post-burn-in BEAST output using “gammaStat” in the R package “ape” (Paradis et al. 2004).

We also tested rates of trait evolution in morphology and habitat and predicted that there would be greater rates in the Madagascar lineages compared to the non-Madagascan lineages. Maximum-likelihood rates of morphological evolution were calculated from the first two PC axes, the individual measurements, and the habitat data using the censored rate test implemented in the program Brownie version 2.1 (O'Meara et al. 2006). This program takes into account the lack of independence between species due to phylogenetic structure as well as time since divergence. The censored rate test was used because this method independently calculates the rate of evolution of groups in different parts of the phylogeny. In a pairwise fashion we compared the rate of diversification in lineages from two different areas (i.e., Madagascar compared to South Africa and Madagascar compared to Australia) by statistically testing whether a single rate model is worse than a multiple rate model whereby each compared group is allowed its own rate of evolution. The maximum-likelihood estimates of the rate of morphological and habitat evolution were compared using the Akaike Information Criterion (AIC) and likelihood ratio test (LRT). Due to small sample sizes, we used the AICc and performed 1000 parametric bootstrapping replications on the LRT. To account for uncertainty in branch lengths and phylogenetic topology this analysis was performed on 500 randomly sampled phylogenies from the BEAST output and results were averaged. Outgroups, fossils, and duplicate species were pruned from the phylogenies.

MORPHOLOGICAL ADAPTATION

Natural history and laboratory observations suggest that archaeids, regardless of the carapace shape, prey upon all types of spiders, rather than specific spider groups (e.g., orb-weavers, hunters, sheet-weavers) (Wood et al. 2012). However, morphological traits could instead be linked to habitat rather than prey type, and correlations between morphology and the environment may suggest adaptation. We examined whether morphology was correlated with habitat using phylogenetic generalized least-squares regression (PGLS) (Martins and Hansen 1997). The PGLS analyses were first performed on the BEAST consensus phylogeny assuming different models of evolution, that is, Brownian (Felsenstein 1985), Pagel's lambda (Pagel 1999), accelerate or decelerate (ACDC; Blomberg et al. 2003), and single-peak Ornstein–Uhlenbeck (Butler and King 2004), which then were compared for best fit using the AIC criteria. Then, assuming the best model, PGLS was performed using 500 randomly sampled phylogenies from the BEAST output and results are reported as averages. PGLS was performed treating the separate morphological PC axes as the response variable and habitat (percentage of time found in vegetation) as the explanatory variable. These analyses were performed twice: for all archaeids, and for only the Madagascan archaeids (with pruning of non-Madagascan archaeids). Analyses were performed using the R packages “ape” and “nlme” (Paradis et al. 2004; Pinheiro et al. 2007).

SPECIES RICHNESS AND ENDEMICITY

Distribution patterns on Madagascar were examined to better understand where species were accumulating. We were specifically interested in examining whether montane areas in Madagascar correspond to distribution patterns, as many archaeids occur at higher elevations. The majority of the species distributions were based on sampling that was performed for over 10 years at over 200 localities throughout Madagascar by the CAS Madagascar Arthropod Inventory Survey. Sampling techniques followed a systematic protocol that used a variety of collecting techniques so as not to bias the study toward certain habitats or localities (Fisher 2005). Distributional data were treated as points per species and were translated into a one-quarter degree square grid that was mapped onto Madagascar using the R packages “raster” (Hijmans 2013) and “sp” (Pebesma and Bivand 2005; Bivand et al. 2013). We then calculated species richness and corrected weighted endemism (CWE) per grid cell. Species richness is a measure of the total number of species per grid cell (Linder 2001). “Weighted endemism” (WE) is similar to species richness except that the totaled per species values are weighted by the total number of grid cells occupied by a species, so that a species that occupies four grid cells will be assigned a value of 0.25 per occupied grid cell. CWE was then calculated by dividing WE (combined for all species in the cell) by the total number of species per grid cell (Crisp et al. 2001). We then performed a linear regression of the dependent variables of species richness and CWE against altitude. The altitude variable was the maximum extracted per grid cell. Species richness and CWE were calculated for 36 of 37 total Madagascan species. The species Eriauchenius pauliani was excluded from the spatial analyses as it has not been collected since the advent of GPS, so that we do not have comparable precision to those of other Eriauchenius.

ANCESTRAL RANGE RECONSTRUCTIONS

An ancestral area reconstruction was performed for two purposes: (1) to examine shifts into new ecogeographic regions, and (2) to compare ecogeographic distributions among close relatives. Biogeographic regions were based on Du Puy & Moat's (P&M) vegetation types (1996), which correspond to the ecogeographic regions in Madagascar (Fig. 6A), and were classified into three areas: rainforests (P&M vegetation types: evergreen humid at low, mid, and montane elevations, and evergreen sclerophyllous woodland); western deciduous forests (P&M vegetation type: deciduous, seasonally dry); southern spiny dry forests (P&M vegetation type: deciduous dry). Ancestor reconstructions were implemented in the R package “BioGeoBEARS” (Matzke 2014), which reproduces the biogeography programs LAGRANGE (Ree and Smith 2008), DIVA (Ronquist 1996), and BayArea (Landis et al. 2013), while also allowing for cladogenesis (e.g., founder events) by introducing an additional parameter “j.” We used the BEAST consensus phylogeny with pruning of the non-Madagascan taxa and duplicate taxa, and used likelihood, Bayesian, and parsimony methods, with and without the “j” parameter. The best model was selected using the AICc criterion and the LRT.

PHYLOGENETIC ANALYSIS AND DIVERGENCE DATING

Our alignment resulted in a total evidence concatenated dataset with 5599 characters, consisting of 126 morphological characters, 658 base pairs (bp) for COI, 328 bp for H3, 2572 bp for 18S, 1915 bp for 28S. The concatenated dataset had 3342 variable sites, and 2795 phylogenetically informative sites. The total evidence phylogeny is shown in Fig. 3 and the phylogenies from the molecular-only and the morphology-only analyses, including a discussion of differences, are in the Supporting Information. We used the total evidence analysis as the best estimate of the phylogenetic relationships because this analysis incorporates multiple lines of evidence. The total evidence analysis recovered a monophyletic Archaeidae, a monophyletic grouping of the extant archaeids, and a monophyletic grouping of the fossil archaeids. Among the extant archaeids, the Australian clade is sister to the Madagascar + Africa clades. The Madagascar clade is paraphyletic with respect to southeastern Africa. With the exception of the fossil taxa, the majority of the branches are well supported with pp greater than 0.95. The monophyletic extant archaeids have branch support of 0.84. We do not think this is problematic in that extant archaeid monophyly is strongly supported by molecular data (pp = 1.0) and morphological data (pp = 1.0).

Regarding the estimation of timing of diversification, there were moderate amounts of rate heterogeneity, meaning that the data are not clock-like: the mean coefficient of variation was 1.041 and the mean ucld.stdev was 0.843. There was not strong evidence for autocorrelation (mean covariance = 0.041). The resulting summary chronogram is presented in Fig. 4. The analysis with COI excluded had weaker branch support at some nodes, but there were no topological conflicts, and the mean divergence date estimations were only slightly younger, and therefore, did not alter our major conclusions (see Supporting Information).

Mean values indicate that archaeid origination began to diversify in the Jurassic-Triassic (Fig. 4, node 1, mean = 205 Ma, 95% CI = 176–237). Diversification of the extant archaeids occurred in the Jurassic with the Australian clade splitting with the Madagascar + South African clade (node 2, mean = 177 Ma, 95% CI = 137–217). The diversification event that gave rise to the Madagascan and southeastern Africa genera occurred 151 Ma (node 3, 95% CI = 115–191), with diversification of the Madagascar “Gracilicollis Group” occurring 116 Ma (node 4, 95% CI = 79–157) and diversification of Madagascan Eriauchenius occurring 108 Ma (node 5, 95% CI = 75–143). Eriauchenius diverged from Afrarchaea a mean of 136 Ma (node 6, 95% CI = 100–173). Afrarchaea began to diversify 82 Ma (node 7, 95% CI = 47–120), and the Australian clade began to diversify 89 Ma (node 8, 95% CI = 45–134). The 95% CI are broad in the majority of estimated divergences.

LINEAGE AND TRAIT DIVERSIFICATION

We ran all subsequent analyses with the individual measurements and the scores from the first two morphological PC axes, which had eigenvalues greater than one, and explain 65.2% and 14.3% of the data (Fig. 5), and combined 79.5%. PC1 describes the elongation and constriction of the “neck,” chelicerae, and legs from the following measurements: carapace constriction, “neck” length, chelicerae length, and femur I length. PC2 describes the angle of carapace tilt.

The gamma statistic was calculated for 500 randomly sampled phylogenies and compared with the recovered critical value from the MCCR test. The majority (92.6%) of the 500 sampled phylogenies had a gamma statistic that was less than the critical value (= −1.787), suggesting that rates of lineage diversification are slowing over time in the Madagascan archaeids (see Fig. S4).

Madagascan archaeids consistently showed a higher rate of trait and habitat diversification compared to the Australian and South African clades (Table 2). These findings are strongly supported (P-value < 0.05) in PC1, carapace length, carapace constriction, “neck” length, and habitat. Although weakly supported, Madagascar also shows a higher rate of diversification in chelicerae length and femur length. The greatest rate difference recovered was for the habitat trait, followed by carapace constriction. Because Madagascan archaeids are a paraphyletic group, having two distinct clades, with the southeastern African Afrarchaea nested within, we performed an additional trait diversification analysis treating Madagascar as two groups. The results of this analysis can be found in the Supporting Information, and did not alter our major conclusions.

MORPHOLOGICAL ADAPTATION

Brownian motion and early burst (ACDC) models were the best fit, with Pagel's lambda having a slightly higher AICc score, and the Ornstein–Uhlenbeck model performed the worst. For this reason PGLS analysis was performed assuming a Brownian motion model, and results are reported as averages of 500 randomly sampled phylogenies (Table 3). There was a strong correlation between PC1 and habitat, confirming that archaeids that live in the vegetation, compared to those that live on the ground, have longer, more constricted “necks,” and longer chelicerae, “heads,” and legs.

SPECIES RICHNESS AND ENDEMICITY

Many archaeid species appeared to be microendemics, with 18 of 36 species occupying two or less adjacent grid cells. Because of the intensive sampling effort through CAS and because local endemics were not collected in more distant areas, we tentatively conclude that many Madagascan archaeid species are true local endemics, rather than being artifacts of biased sampling. Species richness measurements ranged from 1 to 8 (average = 2.37) per grid cell on Madagascar, with the highest values occurring in Ranomafana and Masoala national parks (Fig. 6B). Other areas of high species richness include portions of the eastern rainforests, Parc National Montagne d'Ambre in the far North, and Parc National Namoroka in the West. Calculations of CWE reveal that the highest values are found in the southeast (Fig. 6C).

There was no statistical support for a linear relationship between species richness and altitude (P-value: 0.0973, adjusted R²: 0.0399), although the mid-elevations from 900 to 1500 m have the greatest number of species (Fig. 7B). However, there was support for a positive relationship between CWE and altitude, both with and without an outlier point removed—with outlier: P-value: 0.000479, adjusted R²: 0.2273; without outlier: P-value: 0.00000055, adjusted R²: 0.433 (Fig. 7A).

ANCESTRAL RANGE RECONSTRUCTIONS

The ancestral area reconstruction method using the likelihood dispersal-extinction-cladogenesis (DEC) model (Ree and Smith 2008) was selected as the best model, having the lowest AICc score. The LRT found there was no significant difference (P-value = 1) between DEC and DEC + j (j = the cladogenesis parameter), which makes sense because “j” was estimated to be close to zero (j = 0.0086). For the DEC model, the estimated rate of dispersal was 0.0018 and the estimated rate of extinction was essentially zero. The ancestral reconstruction (Fig. 6D) supports two main findings: (1) there were several range expansions to different ecogeographic regions, including a shift to the western deciduous forests within the “Gracilicollis Group” (nodes 1–2 in Fig. 6D); (2) the estimated rate of dispersal is low suggesting that species have mostly retained the same ecogeographic ranges as their ancestors.