Ice ages and butterflyfishes: Phylogenomics elucidates the ecological and evolutionary history of reef fishes in an endemism hotspot

2.1 Materials

Site location, sampling date, and museum voucher information (where available) for each specimen are outlined in Supporting Information Table S1. All butterflyfish species included in this study and their geographic distribution are listed in Table 1.

As our primary objective is to reconstruct the evolutionary history of butterflyfishes known to occur in the Red Sea and adjacent gulfs or seas, we concentrated our sampling efforts on those species and their closest relatives. Although five major Chaetodontidae lineages were sampled, Chaetodon Clade CH1 (Chaetodon robustus and C. hoefleri, restricted to the Atlantic; Cowman & Bellwood, 2013), and multiple bannerfish genera (Amphichaetodon, Chelmon, Chelmonops, Coradion, Hemitaurichthys, and Johnrandallia) without species represented in the Red Sea were not sampled in this study. Two species of the Prognathodes genus were included to facilitate fossil calibration, but were not included in the biogeographic analyses due to their Atlantic distributions (see below).

In total, we sampled 47 chaetodontid species (35% of the entire family), which includes all regional endemics and wide-ranging species found in the Arabian Peninsula region save Roa jayakari, a rare deepwater species distributed from the Red Sea to coastal India; we were unable to secure a tissue sample as part of this study. Eight of these species have not previously been sampled in phylogenetic studies of the family (Bellwood et al., 2010; Cowman & Bellwood, 2011; Fessler & Westneat, 2007; Hodge et al., 2014). Tissues were preserved in a saturated salt-DMSO solution or 95% ethanol prior to processing. This research was carried out under the general auspices of King Abdullah University of Science and Technology's (KAUST) arrangements for marine research with the Saudi Arabian Coast Guard and the Presidency of Meteorology and Environment. The animal use protocol was approved by KAUST's Biosafety and Ethics Committee (KAUST does not provide specific approval number).

2.2 Phylogenomics approach

We employ the sequence capture method of ultraconserved elements (UCEs) to produce millions of reads in parallel from multiple butterflyfish specimens collected from the Gulf of Aqaba in the west (Red Sea) to the Hawaiian Archipelago in the east (Pacific Ocean, PO). UCEs are a class of highly conserved and abundant nuclear markers distributed throughout the genomes of most organisms (Bejerano, Haussler, & Blanchette, 2004; Siepel et al., 2005; Reneker et al., 2012). These markers do not intersect paralogous genes (Derti, Roth, Church, & Wu, 2006), do not have retro-element insertions (Simons, Pheasant, Makunin, & Mattick, 2006), have a range of variant sites (i.e., evolving on different time scales; Faircloth et al., 2012), and have been used to reconstruct phylogenies across vertebrates (Bejerano et al., 2004; Faircloth et al., 2012; Faircloth, Sorenson, Santini, & Alfaro, 2013; McCormack et al., 2013; Smith et al., 2014; Sun et al., 2014), including fishes at both shallow (Mcgee et al., 2016) and deep (Alfaro et al., 2018; Faircloth et al., 2013; Harrington et al., 2016) phylogenetic scales.

2.3 DNA library preparation and next-generation sequencing

DNA was extracted with DNeasy Blood and Tissue kits (Qiagen, Valencia, CA), which included an RNAse A treatment step. Each extracted sample was visualized by gel electrophoresis to assess DNA quality. Total DNA from each extracted aliquot was quantified using a Qubit dsDNA HS Assay Kit (Invitrogen, Carlsbad, CA), and 1.2 µg of DNA per individual sample was fragmented by sonication to 500 base pairs (bp) using a Covaris S2 sonicator (Covaris Inc, Woburn, MA) and used for UCE library preparation. In brief, we end-repaired, adenylated, and ligated fragmented DNA to Illumina TruSeq-style adapters, which included custom sequence tags to barcode each individual sample (Faircloth & Glenn, 2012). Following an 18-cycle PCR to amplify indexed libraries for enrichment, we created pools by combining 62.5 ng of eight individual libraries. Each pool was concentrated to 147 ng/μl using a vacuum centrifuge. We then followed an established workflow for target enrichment (Gnirke et al. 2009) with modifications specified in Faircloth et al. (2012). Specifically, we enriched each pool, targeting UCE loci and their flanking sequence, using synthetic RNA capture probes (MyBaits, Mycroarray, Inc., Ann Arbor, MI). We combined the enriched, indexed pools at equimolar ratios prior to sequencing. The two final pooled libraries were each run paired-end (150 bp sequencing) on independent lanes of an Illumina HiSeq2000 (v3 reagents) at the KAUST Bioscience Core Lab. Detailed methods of library enrichment, post-enrichment PCR, and validation using relative qPCR may be found at https://ultraconserved.org/#protocols.

2.4 Sequence read quality control, assembly, and UCE identification

We removed adapter contamination and low quality bases with illumiprocessor (Faircloth, 2013), a parallel wrapper to Trimmomatic (Bolger, Lohse, & Usadel, 2014). To assemble the trimmed dataset, we used the PHYLUCE pipeline (version 8ca5884; Faircloth, 2016) with the phyluce_assembly_assemblo_trinity.py wrapper script for Trinity (version 1.5.0; Grabherr et al., 2011). We matched assembled contigs to enriched UCE loci by aligning contigs from each species to our UCE probes using the phyluce_assembly_match_contigs_to_probes.py script with the LASTZ assembler (Harris, 2007). We stored these match results into a SQLite relational database after excluding contigs that matched multiple UCE loci and UCE loci whose probes matched multiple contigs.

We used phyluce_align_seqcap.py to align UCE loci with MAFFT (Katoh & Standley, 2013; Katoh, Misawa, Kuma, & Miyata, 2002). Following alignment, we end- and internally-trimmed alignments with GBLOCKS (Castresana, 2000) to improve phylogenetic inference by removing poorly aligned or highly divergent sites (Talavera & Castresana, 2007). We selected loci that were present in at least 75% of our specimens and concatenated the alignments into a PHYLIP-formatted matrix for phylogenetic analysis. We included previously published UCE data for three species in our alignment to represent Acanthomorpha outgroup lineages and more accurately calibrate the phylogeny (see below).

2.5 Phylogenetic analysis of concatenated UCE data: evaluation of the “evolutionary incubator” and “Pleistocene extirpation” hypotheses

We fully partitioned our concatenated alignment by UCE locus and performed Bayesian analyses of the dataset with ExaBayes (Aberer, Kobert, & Stamatakis, 2014) and two independent runs, sampling every 500 generations. We used the autostopping convergence criteria of an average standard deviation of split frequencies of <5% and visualized the log-likelihood of each chain to ensure convergence in Tracer version 1.6 (Rambaut et al., 2014).

We estimated divergence times using MCMCTREE in the PAML package on the Bayesian consensus topology. We used the likelihood approximation approach following the two-step procedure described by Dos Reis and Yang (2011) by first estimating a mean substitution rate for the entire alignment with BASEML under a strict molecular clock and then using this estimate to set the rgene_prior in MCMCTREE. We used a single, unpartitioned alignment for computational tractability, with an HKY85 model, five categories for the gamma distribution of rate heterogeneity, an rgene_gamma prior for the gamma distribution describing gene rate heterogeneity of (2, 371.0575, 1) and a sigma2_gamma prior of (2, 5, 1). We adopted a calibration strategy that builds on Harrington et al. (2016) by including more proximal acanthomorph outgroups to Chaetodontidae and their immediate relatives. We constrained six nodes on the basis of fossil information using hard lower and soft upper bounds outlined in Supporting Information Figure S1. We assigned a minimum amount of prior weight for ages below the lower bound (1e-200) and 5% prior weight for ages higher than the upper bound. Briefly, we link a series of carangimorph, syngnathiform, holocentroid, and lampridiform fossils to the sequences of acanthomorph outgroup fossils as per Harrington et al. (2016). This resulted in the following outgroup node calibrations: acanthuroids versus all other taxa (lower bound: 54.17 Ma; upper bound: 70.84 Ma); acanthurids versus zanclids (lower bound: 49.0 Ma; upper bound: 62.7 Ma), Naso versus Acanthurus (lower bound: 49.0 Ma; upper bound: 57.22 Ma), Chaetodontidae versus Pomacanthidae (lower bound: 29.62 Ma; upper bound: 59.26 Ma), and the total-group Chaetodon versus Prognathodes (lower bound: 7 Ma; upper bound: 47.5 Ma). Further justification for calibrations is available as Supporting Information (Appendix S1).

2.6 Ancestral biogeographic range estimation: evaluation of the “evolutionary incubator” and “Pleistocene extirpation” hypotheses

We estimated ancestral distribution patterns for chaetodontid lineages using the pruned time-calibrated phylogeny analyzed with the R package BioGeoBEARS (Matzke, 2013), which allows several models of biogeographic evolution to be compared via likelihood inference, and the ability to incorporate a parameter allowing for founder-event speciation. For these analyses, we coded each taxon based on presence/absence in nine discrete geographical areas: Gulf of Aqaba, rest of the Red Sea, Djibouti and Gulf of Aden, Socotra, South Oman, Arabian Gulf, Gulf of Oman and Pakistan, rest of Indian Ocean, and PO. The discrete coding of geographic areas adjacent to the Arabian Peninsula enables a fine-scale investigation of the ancestral biogeography of that region for our taxa of interest. Presence/absence and geographical range data for each taxon were obtained from a combination of DiBattista, Roberts, et al. (2016) and FishBase (Froese & Pauly, 2011). Prognathodes spp. (a Chaetodontidae genus) were not considered in this part of the analysis given that these two taxa are restricted to tropical Atlantic waters.

We constrained our biogeographic analyses to prohibit colonization events between the Red Sea and Indian/PO regions before 5 Ma reflecting the time when a more permanent connection was formed via the Strait of Bab al Mandab (Bailey, 2010). Our BioGeoBEARS analysis evaluated the DEC, DIVALIKE, and BAYAREALIKE models with and without the jump (J) parameter (Matzke, 2013). These models describe biogeographic scenarios where dispersal, extinction, cladogenesis, vicariance, and founder events are differentially invoked to explain present day distributional patterns. In our case, we were interested in whether the range-restricted endemics from the coastal waters of the Arabian Peninsula represent ancient relicts, new colonization events, and/or a source of biodiversity (at some point in the past) for the broader Indo-West Pacific.

2.7 Comparative trait analysis: evaluation of the “ecological trait” hypothesis

In order to determine whether particular species-level traits were associated with the evolution of endemism in this subset of Chaetodontidae species, we fitted a phylogenetic generalized linear model (function “phyloglm” in R package “phylolm” [Ho et al., 2016]) that assumed “regional endemism” (i.e., endemic to the coastal waters of the Arabian Peninsula; DiBattista, Roberts, et al., 2016) as the binomial response variable and a suite of ecological traits as the predictive fixed factors. For model selection, we performed a backward stepwise procedure for PGLM's (function “phylostep” in R package “phylolm” [Ho et al., 2016]), which entailed sequential optimization by removing non-influential fixed-effect terms from the full model based on Akaike information criteria (AIC). Full details on the methods and data sources are provided in Supporting Information Table S2. We also explore interactions among the predictive traits using a regression tree approach (De’ath and Fabricius, 2000; function “rpart” in R package “rpart” [Therneau et al., 2015]).

Among the predictive variables considered were: maximum body size (total length = TL; Allen et al., 1998; Kuiter, 2002), depth range inhabited (Allen et al., 1998), social structure (three categories ordered from low to high sociability: solitary, pair formation, and group formation; Allen et al., 1998; Kuiter, 2002; Yabuta and Berumen, 2013), habitat breadth (estimated as the sum value of all habitat types inhabited: C = coral, R = rocky, D = deep reef, S = sediment, R = rubble, CO = coastal, CA = algal beds; Allen et al., 1998; Kuiter, 2002), and dietary reliance on coral reefs (four categories ordered from low to high reliance: planktivore, benthic invertivore, facultative corallivore, and obligate corallivore; Cole and Pratchett, 2014). We also included the phylogenetic age of species (Myr) as an additional fixed factor to test whether species traits are influenced by time of divergence from sister taxa. For phylogenetic age, we evaluate for each species (regional endemic and widespread) whether we sampled its closest sister species by comparing our phylogeny with those published previously (Cowman & Bellwood, 2011) and other published accounts (Kuiter, 2002). The ecological traits were selected because they are associated with specialization, fitness, and range expansion in butterflyfishes, and thus may help to explain patterns of evolution in fish endemic to the coral reefs of the Arabian Peninsula. We do note this may be an oversimplification given that our categories are coarse and biased toward adult versus larval traits, which are themselves data deficient. Previous work, however, has demonstrated that traits associated with the successful recruitment of reef fish are more important than traits associated with dispersal in determining differentiation between habitats (Gaither et al., 2015; Keith, Woolsey, Madin, Byrne, & Baird, 2015).

3.1 UCE sequences

Reads, contigs, and UCE loci per individual are outlined in Supporting Information Table S3. In summary, we sequenced a total of 153.31 million reads, with a mean of 1.55 million reads per sample from 47 focal taxa (excluding outgroups; also see Table 1). Overall, we assembled a mean of 12,969 contigs (95 CI, min = 10,593, max = 15,345) and 901 UCE loci per sample (95 CI, min = 871, max = 932).

3.2 Phylogenetic reconstruction and timing of divergence: evaluation of the “evolutionary incubator” and “Pleistocene extirpation” hypotheses

Following assembly, alignment, trimming, and filtering out loci that were present in fewer than 75 specimens (for a 75% complete dataset), we retained 971 alignments with a mean length of 515.6 bp. The concatenated supermatrix contained 500,642 bp with 52,680 informative sites and was 83.3% complete based on the proportion of non-gap sequences. The following samples were excluded from further analysis due to the low number of loci recovered: Chaetodon_interruptus1a, Chaetodon_lineolatus1a, Chaetodon_lunula1a, and Chaetodon ulietensis1a (for full details see Supporting Information Table S1); however, tissue replicates were retained for two of the four species listed here (Chaetodon lineolatus and Chaetodon lunula).

Our Bayesian and maximum likelihood analyses produced a fully resolved topology that shared key points of congruence with prior multi-locus studies of butterflyfishes (Bellwood et al., 2010; Cowman & Bellwood, 2011; Fessler & Westneat, 2007; Hodge et al., 2014 ; Hsu et al., 2007; see Supporting Information Figure S2). Although direct comparisons to previous phylogenies are difficult because these are missing many of the regional endemics (e.g., Chaetodon dialeucos, C. gardineri, C. leucopleura, C. nigropunctatus, C. pictus, C. triangulum, Heniochus intermedius), and contain less sequence data and data overlap (e.g., six loci and 73% complete matrix; Hodge et al., 2014), where there was overlap in the data sets the tips of the tree displayed similar topologies (Supporting Information Figure S3). In our case, however, almost every node in the tree was strongly supported (posterior probabilities of 1.0; Figures 1 and 2).

By only considering a single representative sample per species on our chronogram (Figure 2), we found that the majority of Red Sea to Arabian Gulf butterflyfish diverged from their closest relatives in the last five million years (4.17–1.16 Ma), with an average lineage age of 2.39 Ma. In comparison to previous fossil calibrated studies of Chaetodontidae (Cowman & Bellwood, 2011; Hodge et al., 2014), the mean ages and 95% highest posterior density (HPD) estimates are more restricted, but for the most part overlap with previous estimates (Supporting Information Figure S3). In terms of the topology, although our phylogenetic sampling is restricted, it still captures crown nodes and age estimates of major chaetodontid lineages (with the exception of the bannerfish lineage), as well as subclades containing Red Sea to Arabian Gulf species and their most recent common ancestors (Supporting Information Figure S2). Most of the clades included species pairs diverging with non-overlapping distributions dating back 2–1 Ma. This divergence does not appear to coincide with the timing of the emergence of apparent geographic (and geological) barriers such as the Strait of Bab al Mandab (Figures 2 and 3). Regional endemics appear to have diverged earliest from ancestors that gave rise to the clades including Chaetodon larvatus and Chaetodon semilarvatus. At least one entire subclade of CH4 was comprised of regional endemics (C. dialeucos, C. nigropunctatus, and C. mesoleucos). The split between the butterflyfishes (Chaetodon, Prognathodes) and bannerfishes (Heniochus, Forcipiger) was much older, with a mean of 28.7 Ma (95% HPD: 40.0–18.26; Figure 2 and Supporting Information Figure S1), indicating an ancient split between these highly divergent body forms.

3.3 Ancestral range reconstruction: evaluation of the “evolutionary incubator” and “Pleistocene extirpation” hypotheses

Model comparison revealed the DEC+J model as the most likely (LnL = −250.79, AIC weight = 0.76) and the DIVALIKE+J model as the second most likely (LnL = −252.76, AIC weight = 0.11; Table 2 and Figure 4). The importance of the J parameter, which models long-distance or “jump” dispersal, is that ancestral ranges often comprise one area rather than several areas. The addition of the J parameter resulted in a significantly better model fit for DEC models when compared via a likelihood ratio test (LRT: D = 8.67, p = 0.0032).

Under the DEC+J model, Chaetodontidae had their crown origins in the Indo-West Pacific, with subsequent dispersal to include the Arabian Peninsula and lineages leading to the base of Chaetodon and the bannerfish clade (Forcipiger/Heniochus; Figure 4). Within the CH2 clade, diversification was restricted to the PO with subsequent dispersal to the Indian Ocean (Chaetodon madagaskariensis, C. punctatofasciaticus, and C. unimaculatus), and three of the species have dispersed as far as Socotra (Chaetodon guttatissimus, C. kleinii, and C. trifasciatus). Only one species within CH2 diverged in the Gulf of Aden and subsequently colonized the Red Sea (Chaetodon paucifasciatus). The age of C. paucifasciatus is relatively young (1.5 Ma, HPD: 0.8–2.3 Ma), suggesting a similar timeline for its occupation of the Red Sea from the Gulf of Aden.

In the CH3 clade, three species were present in the Red Sea that were also restricted to the Arabian Peninsula (Chaetodon austriacus, C. melapterus, and C. larvatus). In the case of sister pair C. austriacus and C. melapterus, the reconstruction suggests that speciation occurred by vicariance within the Red Sea. Although posterior probabilities make the details of this split equivocal, the most likely scenario is a split between the Gulf of Aqaba and the Red Sea, where C. austriacus subsequently recolonized the entire Red Sea but C. melapterus expanded out to the Gulf of Aden, Arabian Sea, and Arabian Gulf, but no further. The extended history of the clade, although not completely sampled (see Supporting Information Figure S2), suggests that a widespread ancestor expanded into the Red Sea with subsequent vicariance between the PO, Indian Ocean, and Arabian Peninsula sites. Indeed, C. larvatus appears to originate in Djibouti and the Gulf of Aden followed by dispersal into the Red Sea and South Oman. Chaetodon trifascialis, on the other hand, maintained connections across the Indo-West Pacific with subsequent range expansion into the Red Sea.

The CH4 clade has been the most successful in terms of butterflyfish colonizing the Red Sea. Eight extant species from CH4 are distributed in at least some parts of the Red Sea (Chaetodon auriga, C. fasciatus, C. leucopleura, C. lineolatus, C. melannotus, C. mesoleucos, C. pictus, and C. semilarvatus), four of which are restricted to the Arabian Peninsula (C. fasciatus, C. mesoleucos, C. pictus, and C. semilarvatus). Moreover, the reconstruction identified a mix of origin states for CH4 species found in the Red Sea. Both C. fasciatus and C. leucopleura have their origins within the Red Sea, whereas C. lineolatus and C. mesoleucos have their origins at Socotra. The origins of C. semilarvatus are placed in South Oman, whereas the origins of C. pictus are placed in the Gulf of Oman. With the exception of C. lineolatus, a widespread Indo-West Pacific species, all CH4 lineages have origins in the Arabian Peninsula region and Indian Ocean, and subsequent dispersal was limited from these sites. Chaetodon lineolatus appears to be the only species in CH4 to originate in the Arabian Peninsula and then disperse across the broader Indo-West Pacific. However, unsampled taxa from this clade are more widely distributed across the Indian and POs (Supporting Information Figure S2).

Three taxa of the bannerfish clade are also present in the Red Sea (Heniochus diphreutes, H. intermedius, Forcipiger flavissimus), with H. intermedius considered a Red Sea to Gulf of Aden endemic. Despite these taxa only being representative of a small proportion of the entire bannerfish clade, the reconstruction suggests a widespread ancestor that diverged in the Arabian Peninsula region (H. intermedius) with subsequent (successful) colonization of the broader Indo-West Pacific (H. diphreutes and F. flavissimus).

3.4 Correlational trait analysis: evaluation of the “ecological trait” hypothesis

Based on the best-fit PGLM, depth range and phylogenetic age were negatively correlated with endemism, with depth range being a stronger predictor than phylogenetic age (Table 3, Figures 5 and 6). Exploring these relationships using a regression tree approach reveals that the effect of phylogenetic age is dependent on depth range. Endemic species from the Arabian Peninsula region are therefore more likely to be younger than widespread ones, but only for those species with depth ranges extending to mesophotic reefs (depth range >27 m; Figures 5 and 6). Endemism was not correlated with any of the other factors in the analysis for the butterflyfishes considered here (Supporting Information Tables S2 and S4).

4.1 Evaluation of the “evolutionary incubator” hypothesis

The Red Sea, Gulf of Aden, Arabian Sea, and Arabian Gulf are all peripheral to the broader Indo-West Pacific biogeographic region and potentially produce/contribute new reef fish species to the center (see Bowen et al., 2013; Hodge et al., 2014). Temporally, the Red Sea to Arabian Gulf butterflyfish assemblage (17 species in total) is made up of recently diverged lineages, with ages ranging from 4.17 Ma (F. flavissimus) to 1.16 Ma (C. austriacus/C. melapterus split). In a few cases, the Red Sea to Gulf of Aden endemics appear to have diverged as the earliest lineage of that clade (e.g., C. larvatus and C. semilarvatus; Figures 2 and 3). Indeed, the “oldest” endemic butterflyfish lineage in our study, C. larvatus (2.86 Ma, 4.3–1.6 Ma 95% HPD), appeared in the late Pliocene, and diverged from an Indo-West Pacific lineage that later gave rise to species allopatric between the two ocean basins (C. triangulum in the Indian Ocean and C. baronessa in the PO). The ancestral range reconstruction of these Arabian Peninsula endemics demonstrates consistent colonization routes to the Red Sea and Arabian Sea via the Indian Ocean from the east (Figure 4), but with few examples of reciprocal expansion from the Arabian Peninsula back to the Indian Ocean and PO. For example, both C. larvatus and C. semilarvatus appear to have historically diverged in Djibouti/Gulf of Aden and South Oman, respectively, successfully colonized the Red Sea, but not established further south and east based on present day distributions. Similar reconstruction results were obtained for the regional endemic C. pictus (Red Sea to Gulf of Oman), which showed apparent historical divergence in the Gulf of Oman and only recent colonization of the southern limits of the Red Sea.

Other endemics appear to have historically diverged within the Red Sea (C. fasciatus) or adjacent Djibouti and Gulf of Aden (C. paucifasciatus) but not colonized any further to the southeast. Although equivocal based on the probabilistic uncertainty of nodes in the ancestral range reconstruction of the most likely model (DEC+J), there are a number of competing explanations for how C. austriacus and C. melapterus diverged from each other within the coastal waters of the Arabian Peninsula (also see Waldrop et al. 2016), particularly since C. melapterus is the only species in this complex present in the Arabian Gulf. The most likely explanation is based on present day distributions (Figure 3c): C. austriacus is largely restricted to the northern and central Red Sea (with rare records in the southern Red Sea and outside of the Red Sea), whereas C. melapterus is most abundant within or adjacent to the Arabian Gulf (with rare records in the southern Red Sea)—these bodies of water show opposite trends in terms of productivity, sea surface temperature, and temporal patterns of environmental variation (Pous, Lazure, & Carton, 2015; Raitsos et al., 2013). These environmental conditions are additionally significantly different from the rest of the Indian Ocean, and thus, the unique conditions in the Red Sea and Arabian Gulf may help explain how endemics evolved, or at least, concentrated and persisted in these peripheral locations.

Despite a lack of supporting evidence for the evolutionary incubator hypothesis, a clear pattern emerges that the unique environmental conditions in these peripheral seas may have contributed to the formation of endemic species as outlined above. For example, some butterflyfish subclades are comprised entirely of regional endemics (e.g., C. dialeucos, C. mesoleucos, and C. nigropunctatus), which provides further evidence that coral reef habitat surrounding the Arabian Peninsula may have generated a number of new taxa. Moreover, C. dialeucos, an Omani species, shows geographical divergence with the remaining taxa in its group (Figure 3), which all went on to colonize the Red Sea and the Arabian Gulf and must have therefore encountered contrasting environments at the western and eastern margins of their range. The shallow Arabian Gulf started to fill with seawater approximately 14,000 years ago after being dry prior to that during the last glacial maximum (Lambeck, 1996), suggesting that it was seeded by successive waves of colonization from coastal Oman. The same process would have been ongoing at the western margin of the C. dialeucos range, except that the conditions encountered in the Red Sea would have contrasted to those in the Arabian Gulf (DiBattista, Choat, et al., 2016). So, while there is some evidence to suggest vicariance at the scale of the Arabian Peninsula (i.e., diversification of most taxa occurred in the Plio-Pleistocene), a stronger scenario is that natural selection driven by the major differences in environment and habitat within the area probably played an important role in the formation of endemic species assemblages (e.g., Gaither et al., 2015). Thus, even though the distribution of some of the butterflyfishes considered in the present study does stop abruptly at the entrance of the Strait of Hormuz (Chaetodon collare, C. pictus, and C. gardneri), it does not support the argument for geographically driven allopatry. Indeed, all of these species have a different distributional response near the other end of their distribution at the Strait of Bab al Mandab, which includes stopping before the Straits or extending through the Straits into the southern Red Sea (Figure 3). The alternative is that the incumbent widespread butterflyfish may have restricted the Red Sea to Arabian Gulf endemics from expanding further via competitive exclusion.

The current environment of the Red Sea is spatially structured with major contrasts in the cool oligotrophic waters of the northern region compared to the much higher temperatures and productivity of the southern region (i.e., Farasan Islands in Saudi Arabia to the east and Dhalak Archipelago in Eritrea to the west) (Racault et al., 2015; Raitsos et al., 2013). This shift in environmental conditions is most clearly demonstrated in the differences in life history traits associated with reef fish species that occur in both areas, but is also seen in abundance estimates across these gradients (DiBattista, Roberts, et al., 2016; Roberts et al., 2016). Such putative selection gradients are most evident in corals, which show signatures of local adaptation to divergent environmental conditions (D'Angelo et al., 2015).

4.2 Evaluation of the “Pleistocene extirpation” hypothesis

The second hypothesis that we tested in this study was the Pleistocene extirpation hypothesis, which predicts that all Red Sea fauna were eliminated during the last glacial maxima (~18 ka) and were only re-populated via recent colonization events (see Biton et al., 2008). The number of species diverging at early stages in the Pleistocene disputes the argument that Red Sea fauna did not survive complete closure or restriction of water flow at the Strait of Bab al Mandab (Figure 2). Although it clearly does not coincide with a single vicariance event given the variability in the splitting dates between closely related species (Figure 3; see Michonneau, 2015 for invertebrate examples) and ancestral range reconstruction favoring +J parameter models (i.e., founder events between non-adjacent ocean regions; see Table 2), glaciations likely played a role in their separation. Moreover, even though almost all sister species have small areas of overlap at their range edge, which is usually associated with allopatric speciation, in our case these do not coincide with geographical boundaries (i.e., vicariant chokepoints) such as the Strait of Bab al Mandab (see Figure 3; Lambeck, 1996; DiBattista, Choat, et al., 2016). In fact, the non-congruent age and distribution of the endemic species indicate a series of variable events, which may reflect localized patterns of habitat and environmental change as outlined in the previous Discussion section. The best example is the relatively young clade of Arabian Peninsula endemics: C. dialeucos, C. nigropunctatus, and C. mesoleucos (crown node age 2.0 Ma; 2.9–1.2 Ma 95% HPD). This group appears to have been influenced by boundaries presented by the Omani coastline across areas where there are known changes in the upwelling regime (McIlwain, Claereboudt, Al-Oufi, Zaki, & Goddard, 2005; Shi, Morrison, Bohm, & Manghnani, 2000). This is in sharp contrast to the Indo-West Pacific parrotfishes, where present day species boundaries support the notion of allopatric divergence (Choat, Klanten, Herwerden, Robertson, & Clements, 2012), and endemics appear to have diverged into one or more subsequent endemics (i.e., secondary endemism; Rotondo, Springer, Scott, & Schlanger, 1981) based on sympatrically distributed sister-species pairs (highlighted in Choat et al., 2012). Moreover, Red Sea endemics from most other families of reef fish appear to have equal proportions of allopatrically and sympatrically distributed sister species (Hodge et al., 2014), which is not the case for the butterflyfishes.

The diversification of these butterflyfishes occurred at a time when the coral assemblages of the world's reefs underwent a major change in coral composition and growth forms. The global proportion of staghorn coral occurrences in coral assemblages persisted throughout most of the Cenozoic but increased substantially during the Pliocene and especially the Quaternary (Renema et al. 2016). Indeed, the rapidly growing branching acroporid corals offered different structural components in terms of shelter and feeding/foraging modes when compared to massive corals such as poritids that dominated Miocene reefs more than 5 Ma. Thus, the chaetodontids of the Arabian Peninsula (particularly the corallivorous species) were exposed to a much more dynamic environment than the widespread Indo-West Pacific species (Coles, 2003) because of their close association with sensitive coral genera that proliferated in the region.

4.3 Evaluation of the “ecological trait” hypothesis

The third hypothesis that we test here is whether ecological traits are linked to the evolution of endemism among butterflyfishes in the Red Sea to Arabian Gulf. We found a negative, significant relationship between endemism and depth range and, to a lesser extent, phylogenetic age for these butterflyfishes (Figures 5 and 6). The relationship between a narrow versus broad depth range and endemism supports the view that endemic species tend to be more specialized to local resources than widespread species (Hawkins, Roberts, & Clark, 2000). The majority of regional endemics in this study had depth ranges that did not extend deeper than 25 m (Figure 6), despite the availability of light dependent coral habitat extending beyond that (Kahng et al., 2010). The broad range of ages represented by these shallow water specialists suggests that adaptation to shallow reefs occurred multiple times across a relatively wide time frame (i.e., 1.3–3.3 Myr). On the other hand, speciation of endemics with a preference for deep reefs seems to be a recent phenomenon, as deeper depth ranges were strongly associated with young age (<1.75 Myr; Figure 6).

4.4 Comments on incomplete sampling and biogeographic biases

The goal of this study was to reconstruct the evolutionary history of Red Sea to Arabian Gulf butterflyfishes. As is the case with all phylogenetic and biogeographic reconstructions, our results have to be interpreted in light of the taxa that are not sampled, both extant and extinct. Indeed, the inclusion of missing taxa has the potential to alter lineage relationships and their age estimates, whereas their geographic distribution may alter the most likely biogeographic scenarios reconstructed across the tree (see discussion in Cowman & Bellwood, 2013). Here, we were able to sample all Red Sea to Arabian Gulf butterflyfishes (save one species, R. jayakari), and their close relatives from the Indian Ocean and PO, across four major chaetodontid lineages (Supporting Information Figure S2). From a temporal perspective, the topology and ages estimated for the genomic scale UCE data overlap with previous studies (Supporting Information Figures S2 and S3). Moreover, our sampling of eight species that have not previously been included in phylogenetic studies of the Chaetodontidae family means that for 13 out of the 17 Arabian Peninsular species, we are confident that we have sampled their direct sister lineage. Two of the outstanding three species (C. melannotus, C. trifascialis) are wide-ranging Indo-West Pacific taxa that are reconstructed to have dispersed to the Arabian Peninsula (Figure 4). The most likely sister species of C. melannotus is C. ocellicaudus (Kuiter, 2002; also see Supporting Information Figure S2), a west Pacific species not sampled in our dataset. In the case of C. trifascialis, it is placed as the sister lineage for a subclade of CH3 containing 10 species distributed across the Indian Ocean and PO, of which we sampled four species (Supporting Information Figure S2; Cowman & Bellwood, 2011). The final outstanding species, C. leucopleura, is placed as a sister species to C. gardineri. Both species have not previously been sampled in phylogenetic studies, but are recognized to be closely related to a third species, Chaetodon selene (widespread in the west Pacific, Kuiter, 2002), which was not sampled in our UCE dataset. In each of these three cases, and more broadly across the family, the inclusion of unsampled species would increase the influence of the Indian Ocean and PO in the ancestral estimation of biogeographic ranges. As such, it would act to strengthen our conclusion that even though the Red Sea and adjacent gulfs and seas have been important for the generation of endemic species, they have had little contribution to the wider Indo-West Pacific diversity of butterflyfishes.