2.1 Taxon sampling

We sampled 62 individuals overall for the three species within the complex using both specimens from Bond (2012) and new records (Figure 1; see Table S1 for locality information). Aptostichus barackobamai was collected across its geographic range in northern California for a total of 21 samples, and A. icenoglei was collected throughout its range in southern California for a total of 40 samples. Only one specimen of A. isabella was included in this study due to collecting constraints (i.e., only one individual of this species has ever been collected and a burrow has not yet been found containing this species; Bond, 2012).

2.2 Sequence capture

Data for ultraconserved elements were produced following the methods described in Faircloth et al. (2012) with subsequent modifications in Starrett et al. (2017), Hedin et al. (2018), and Kulkarni et al. (2020). We extracted genomic DNA from leg tissue for A. barackobamai and A. icenoglei individuals using the Blood and Tissue DNeasy kit (Qiagen) following the manufacturer's protocol. The lone A. isabella individual, a museum voucher kept in 80% ethanol and stored at room temperature, had DNA extracted from leg tissue following the “MMYT protocol” from Tin et al. (2014) with modifications in Derkarabetian et al. (2019). DNA quantification and quality check were performed using Qubit 3.0 Fluorometer (Life Technologies) and agarose gel, respectively. Excluding A. isabella, 250 ng of DNA was sonicated into fragments ranging from 200 to 1000 bp using an ultrasonicator (Covaris E220). UCE libraries were generated with the KAPA Hyperprep Kit (Roche) with universal adapters and iTru5/7 barcodes (Glenn et al., 2019; BadDNA@UGA) with slight modifications on a few steps for A. isabella (for details see Derkarabetian et al., 2019). Libraries were hybridized at 60°C for 24 h to the Spider probe set (Kulkarni et al., 2020) following the version 4 chemistry protocol (Arbor Biosciences). Hybridization-enriched library pools were sequenced with 150 bp paired-end reads on the HiSeq4K at the University of California Davis DNA Technologies Core. Additional individuals were sent to Rapid Genomics (Florida) for library preparation and sequencing.

Sequence processing and analyses were performed on the Farm Community HPC at the University of California, Davis. Reads were filtered and trimmed using Illumiprocessor (Faircloth, 2013) and Trimmomatic (Bolger et al., 2014) in the Phyluce 1.7.1 pipeline (Faircloth, 2015). De novo assemblies with the cleaned paired-end and single-end reads were performed using SPAdes v. 3.14.1 with the isolate option (Prjibelski et al., 2020). Scaffolds were matched with 65% identity and 65% coverage to the modified probe list from Maddison et al. (2020), which is a blend of the Arachnid (Faircloth, 2017; Starrett et al., 2017) and Spider (Kulkarni et al., 2020) probe sets. MAFFT (Katoh & Standley, 2013) was used to align individual locus datasets, and alignments with locus occupancy (i.e., completeness) minimums of 50%, 75%, and 90% were obtained. Alignment masking was performed with TrimAl v.1.2 (Capella-Gutierrez et al., 2009) using default settings.

SNP datasets were generated for A. icenoglei only with Phyluce from the 50%, 75%, and 90% minimum occupancy loci. Reads were mapped against corresponding scaffolds with BWA (Li & Durbin, 2009), implemented in Phyluce, and phased alignments were generated for each minimum locus completeness set. Phased alignments were screened for SNPs, with five sets of single random SNP per locus generated to test for SNP set sensitivity.

2.3 Phylogenetic & biogeographic analyses

Phylogenies were estimated for three different data sets (50, 75, and 90 percent locus completeness; Figure 2 and Figures S1 and S2) with a maximum likelihood inference using IQ-TREE v2.1.2 (Minh et al., 2020). Model selection was performed by ModelFinder (Kalyaanamoorthy et al., 2017), which is implemented in IQ-TREE, and support values were inferred from 1000 replicates of ultrafast bootstrapping (Hoang et al., 2018). Our phylogenies were visualized in FigTree v1.4.1 with midpoint rooting (midpoint rooting produces a result consistent in other analyses; Bond, 2012) and compared to assess congruence among clades. We also conducted two coalescent-based analyses for the 75p and 90p data sets. Gene trees for each locus were constructed using RAxML v8.0.12 for each data set and used to generate a coalescent-based tree with ASTRAL-III (Zhang et al., 2018). Multispecies coalescent (MSC) bootstrapping was run with ASTRAL v.5.7.4 and 100 pseudoreplicates (Simmons et al., 2019). For downstream analyses, we employed the ML phylogeny based on the largest amount of taxon coverage and with the most robust support values (i.e., the phylogeny with 90 percent minimum locus completeness).

Biogeographic analyses were generated using Reconstruct Ancestral State in Phylogenies (RASP; Yu et al., 2015) with dispersal constraints (i.e., dispersal multipliers set to 0.01 for adjacent areas and 0.0001 for non-adjacent areas) to account for their limited dispersal capacity and using our 90p consensus tree from IQ-TREE. Model testing was conducted using the R package BioGeoBEARS (Matzke, 2014), implemented in RASP, and the best-fit model was chosen based on the weighted AICc scores (Figure 3). The distribution range of this complex was divided into seven areas: (A) lower San Diego county; (B) upper San Diego county/Santa Ana Mtns; (C) Transverse Ranges (San Gabriel & San Bernardino Mtns); (D) southern Sierras; (E) northern rim of Central Valley; (F) Sutter Buttes; and (G) Northern Coast Ranges.

2.4 Cohesion species delimitation

To assess species boundaries within A. barackobamai and A. icenoglei, we employed the methodological framework for delimiting cohesion species from Bond and Stockman (2008) that evaluates two cohesion mechanisms: genetic exchangeability and ecological interchangeability. We used our 90p topology as the baseline evolutionary framework for establishing the “basal starting point” to identify potential separately evolving lineages (for details see flowchart in Bond & Stockman, 2008) within A. icenoglei and A. barackobamai. Due to the paraphyletic grade of lineages with respect to geography (i.e., northern Coast Range species did not form a clade) within A. barackobamai, we designated all the individuals as part of one evolving lineage that was not tested further for genetic and ecological exchangeability. For A. icenoglei, we also used our topology from the MSC tree resampling (Figure S6) as additional guidance for establishing lineage designations, which resulted in 3 lineages: North, Central, and South (see Figure 2). We evaluated the distributions of these lineages as well as performed morphological and genetic clustering analyses to assess the potential for gene flow. Genetic exchangeability was rejected if any allopatric lineage forms an apparently separate clustering pattern from other lineages, or if any parapatric lineage has a separate clustering pattern and an obvious barrier to gene flow.

For morphological data, we quantified 25 continuous character measurements for 30 males (10 males from each lineage; Table S2). All measurements were recorded in millimeters and were quantified with a Leica M165C stereomicroscope using the Leica Application Suite software and a digital camera. Measurements were transformed to log-normal values, and a principal component analysis was conducted using the prcomp function in the R package stats (R Core Team, 2022) and visualized in ggplot2 (Wickham, 2016), following Hamilton et al. (2016).

We conducted two genetic clustering analyses. Variational AutoEncoder (VAE), an unsupervised machine learning approach derived from Bayesian probability theory, was used to visualize clustering of these lineages (Figure 4; for details see Derkarabetian et al., 2019). This class of neural networks takes large-scale SNP data as input and compresses this high-dimensional data through several encoding layers into two-dimensional latent variables, which is subsequently reconstructed by uncompressing the latent variables through several decoding layers (Derkarabetian et al., 2019). SNP datasets were converted to one-hot encoding, which converts categorical data into numerical data as needed for certain machine learning algorithms and used as input for VAE analyses. Three replicates per random SNP set were conducted for each dataset (15 total replicates per dataset), and the replicate for each dataset with the least amount of loss during decoding was used for visualization (Figure 4). CLADES, a supervised machine learning approach, was used to further test species hypotheses (Pei et al., 2018). A 90% minimum locus completeness data set with all A. icenoglei individuals was the input data for CLADES. The delimitation analysis was performed using a training model of genetic characteristics of species generated from a short-range endemic arachnid genus (Metanonychus) that has similar natural history characteristics to mygalomorph spiders (Metano_CLADES model from Derkarabetian et al., 2022).

Niche-based species distribution modeling (SDMs) with measures of SDM overlaps for each lineage were used as a proxy for ecological interchangeability, with ecological interchangeability rejected if both the niche equivalency and niche similarity tests are more different than expected by chance (i.e., niche divergence). Current climate data for 19 bioclimatic variables averaged from 1970 to 2000 were downloaded from WorldClim version 2.1 at 30 arc-second spatial resolution (https://www.worldclim.org/data/worldclim21.html; Fick & Hijmans, 2017). Climate layers were cropped to encompass the geographic area of interest and converted to a raster stack using R packages raster (Hijmans, 2015) and rgdal (Bivand et al., 2019). Highly correlated variables with a Pearson correlation coefficient > .80, estimated using the R package ENMTools (Warren et al., 2021), were removed. The remaining bioclimatic variables (see Table S3) were used in conjunction with occurrence records from the current study as well as records from Bond (2012) that could confidently be assigned to a lineage to generate SDMs, with duplicate records deleted prior to SDM construction. The R package ENMeval (Kass et al., 2021) was used to estimate the SDM for each lineage by implementing MaxEnt (Phillips & Dudík, 2008), which is a machine learning program that uses a maximum entropy algorithm. Multiple points within a 30 arc-second grid cell were removed (i.e., only retaining one record per grid cell) by ENMeval during the modeling step to reduce potential for spatial autocorrelation. To limit the likelihood of overfitting while also accounting for goodness of fit, multiple feature classes and regularization multipliers were chosen to generate a total of 30 models (see Tables S4–S7 for model parameters and stats). Model selection was based on AICc, with the best model having a delta AICc of zero and was subsequently used in downstream analyses (Figure 5).

Statistical comparisons of SDMs for each sister lineage comparison were conducted with niche overlap, niche equivalency, and niche similarity tests in ENMTools (Warren et al., 2008, 2010). We used the Schoener's D statistic (Schoener, 1968) to calculate the niche overlap for each lineage comparison, which ranges from 0 (no overlap) to 1 (complete overlap). We carried out two tests, niche equivalency and niche similarity, to evaluate the significance of niche overlap with a randomization procedure (Warren et al., 2008, 2010). The niche equivalency test, a one-tailed test, assesses whether the two niches being compared are identical or not. If the observed niche overlap value is significantly lower than the null distribution of randomized D values, then the niches are not identical (i.e., not equivalent; Figure S7). Considering the limitation of relying only on occurrence records for the niche equivalency test (Warren et al., 2008), we also employed the niche similarity test, a two-tailed test, to assess whether niche overlap between lineages relative to the niche spaces available to those lineages is more similar or different than expected by chance (niche conservatism or niche divergence, respectively; Figures S8 and S9). We estimated three potential background regions for each lineage in ArcGIS Pro v2.8 (ESRI): (1) minimum area polygons based on occurrence points; (2) minimum area polygons based on SDM raster grid cells with a habitat suitability score threshold greater than 0.5 (i.e., a polygon was generated around every grid cell with a habitat suitability score greater than 0.5); and (3) minimum area polygons based on SDM raster grid cells with a habitat suitability score threshold greater than 0.75 (see Figures S10–S12 for reference).

3.1 UCE stats

The UCE data are summarized in Table 1. Across all individuals, contigs that matched to the probes had a mean length of 1010 bp, with an average of 545 contigs over 1 kb per individual. After aligning, filtering, and trimming these UCE contigs, we had three data matrices with varying minimum locus completeness percentages: 50p containing 1336 loci with 1,101,054 total bp, 75p containing 835 loci with 692,091 total bp, and 90p containing 242 loci with 204,134 total bp. For each A. icenoglei SNP data set, there were 1120 SNPs, 668 SNPs, and 195 SNPS for the 50p, 75p, and 90p respectively.

3.2 Phylogenetic and biogeographic analyses

All estimated phylogenies fully supported (i.e., 100 for IQ-TREE or one for ASTRAL analyses) species level divergence among the three previously delineated morphological species within this sibling complex (see Figure 2 and Figures S1–S6). Also, all A. icenoglei lineage divergences (North, Central, and South) were highly, if not fully, supported across the majority of phylogenetic trees (i.e., all IQ-TREE and ASTRAL analyses); however, despite recovering North and South lineages as monophyletic and highly supported (i.e., >90) in both the 75p and 90p MSC bootstrapping analyses, there was uncertainty in placement of several Central lineage individuals with both analyses (Figures S5 and S6).

RASP analysis (Figure 3) inferred an unresolved ancestral range for the ancestor of the complex, with only .1968 probability of a relatively widespread ancestor along the southern Sierras and in the Transverse Ranges that then dispersed to the north with a vicariance event, splitting the ancestor of A. barackobamai from the ancestor of A. isabella + A. icenoglei. Within A. barackobamai, there were dispersal events northeastward along the northern Coast Ranges to the northern rim of the Central Valley/Sierra Nevada, and vicariance events splitting the northern Coast Ranges populations from northeastern populations as well as the Sutter Buttes population from the northern Sierra Nevada population. The split between the ancestor of A. isabella and the ancestor of A. icenoglei, with a .25 probability, was potentially the result of dispersal further south and a vicariance event. For the A. icenoglei lineages, the most likely scenario involved dispersal to the south toward the Peninsular Ranges with vicariance splitting the North from Central+South lineages and subsequent vicariance splitting Central from South.

3.3 Cohesion species delimitation

Table 2 summarizes results for each genetic exchangeability analysis. Geographic distribution assessments for each lineage comparison were considered parapatric. One comparison had no obvious barrier to gene flow (i.e., Central and South lineages), whereas the other comparison of North and Central+South lineages had an unlikely chance of gene flow occurring due to the LA Basin acting as a geographic barrier (see Figure 1). Three clustering analyses, one with morphological data and two with molecular data, were also used to inform the possibility of gene flow. PCA analysis of the quantitative morphological measurements reveal no distinct clustering for any of the lineages (Figure 4). Similarly, the CLADES analysis with the Metano_CLADES training model identified one species. In contrast, the VAE analysis for 50p indicates three very distinct clusters corresponding to each lineage for both the mean and standard deviation (Figure S13); however, although VAE analyses for 75p and 90p show separation between the lineages for the mean, there is a small amount of overlap for the standard deviation between Central and South lineages (Figure 4).

Table 2 summarizes results for each ecological interchangeability analysis. Niche equivalency was rejected for both lineage comparisons, indicating that their niches are not identical (Figure S7). Niche similarity test results were different depending on the background region selected. Central occurrence points compared with the South background, determined by a minimum bounding polygon connecting its occurrence points, were not significantly different; however, the reciprocal comparison of South occurrence points to Central background was significantly more similar than expected by chance (i.e., niche conservatism; Figure S8). Central occurrences compared with the South background, determined by minimum bounding polygons based on raster grid cells with either habitat suitability scores >0.5 or >0.75, and vice versa indicated niche conservatism (i.e., more similar than expected compared with the null distribution; Figure S8). When comparing the Central+South occurrence records to the minimum bounding polygon connecting occurrence points defining the background region of North, the results show no significant difference; yet, the reciprocal comparison is significantly more similar than expected (Figure S9). All comparisons of North versus Central+South and vice versa suggest niche conservatism when background regions are defined by minimum bounding polygons based on raster grid cells with either habitat suitability scores >0.5 or >0.75 (Figure S9).

4.1 Speciation and phylogeography

Spiders in the A. icenoglei complex, like most mygalomorphs, have limited dispersal capabilities and relatively long generation times (Bond, 2012; Harvey et al., 2018; Hedin et al., 2013; Hendrixson et al., 2013), which contributes to their tendency to have population structure at relatively small spatial scales. The molecular data show genetic divergence across the A. icenoglei complex and within A. icenoglei populations, thus populations have likely been isolated from gene flow for a long period of time, indicating the increased potential for speciation (Barraclough, 2019). The three nominal species (A. barackobamai, A. isabella, and A. icenoglei) are distributed across different regions of the CAFP and have been delimited based on distinct morphological differences in secondary mating structures (clasper morphology; Bond, 2012), providing evidence that gene flow has not occurred between them for a long period of time. Within A. icenoglei lineages, our VAE analyses with 75p and 90p datasets support the lack of gene flow between the North and Central+South lineages over an extended time period.

All species within the complex, with potentially the exception of A. isabella, seem to have similar microhabitat requirements (e.g., north-facing shady slopes) despite their occurrence in different ecoregions, similar to other mygalomorph taxa in the CAFP (Hedin et al., 2013; Hedin & Carlson, 2011; Leavitt et al., 2015). Our niche similarity tests show evidence of niche conservation within A. icenoglei lineages, with the caveat that various analyses yielded different results when minimum bounding polygons versus raster grid cell thresholds parameters were considered. Many studies have used the niche similarity test to evaluate overlap in niche space (Hendrixson et al., 2013; McCormack et al., 2009; Newton et al., 2020; Starrett et al., 2018; Warren et al., 2008), yet very few are explicit about the background region they chose to incorporate into the analysis (McCormack et al., 2009; Newton et al., 2020; Starrett et al., 2018). In addition, as far as we are aware, no other study other than Newton et al. (2020) has explicitly tested multiple background regions to evaluate the impact background region choice has on the analysis. Our background region choices for the current study were chosen based on previous studies (minimum bounding polygon; McCormack et al., 2009; Starrett et al., 2018) and polygons reflecting raster grid cells with habitat suitability score thresholds (> 0.5 and > 0.75) that better reflect the suitable habitat space available (i.e., not including large gaps of uninhabitable areas that are included in the minimum bounding polygon). The minimum bounding polygon yielded conflicting results for both North versus Central+South comparisons and Central versus South comparisons, which is most likely attributed to the aforementioned uninhabitable areas included in the analysis that potentially obscures a signal of niche conservatism. Although we attempted to incorporate a more biologically realistic background region, it is possible that our habitat suitability thresholds slightly inflated the inference of niche conservatism and thus may warrant additional future testing and evaluation.

Niche conservatism, in conjunction with restricted gene flow, suggests that speciation scenarios in which vicariant events separate populations with subsequent reproductive isolation through genetic drift, as opposed to ecological differentiation, may apply across the A. icenoglei complex. This pattern is also supported by our biogeographic analysis, but caution should be used when interpreting these results considering our ultrametric tree was not dated and, thus, cannot pinpoint with certainty the geological/climatic events that potentially influenced these splits. First, a vicariant event (after range expansion; Figure 4) is inferred for the split of the ancestor of A. barackobamai and ancestor of A. isabella + A. icenoglei. This phylogeographic break potentially coincides with uplift of the Transverse Ranges approximately 5 mya (Norris & Webb, 1990), which likely cut off the potential for gene flow, and has been hypothesized for other CAFP taxa (Alexander & Burns, 2006; Calsbeek et al., 2003; Feldman & Spicer, 2006; Reilly et al., 2015; Rissler et al., 2006). Second, the split between A. isabella and A. icenoglei possibly occurred due to vicariance. This split could be attributed to the Tehachapi Mountains acting as a barrier to dispersal, which has also been inferred for other taxa (Calsbeek et al., 2003; Chatzimanolis & Caterino, 2007; Rissler et al., 2006). Third, vicariance was inferred for the split between the North lineage and Central+South lineages, which could be associated with periodic inundations of the LA Basin (Jacobs et al., 2004) that might have resulted in habitat fragmentation, also hypothesized for the mahogany Jerusalem cricket (Stenopelmatus “mahogani”; Vandergast et al., 2006) and stream-dwelling frogs (Pseudacris cadaverina; Phillipsen & Metcalf, 2009).

4.2 Species limits within Aptostichus icenoglei and taxonomic implications

Although our integrative approach considered multiple independent lines of evidence, our conflicting results circle back to the unavoidably subjective question of how much weight should be given to genetic divergence versus morphological/ecological divergence (or lack thereof) when delimiting species with extreme population structuring. Should we elevate genetically diverged lineages to species status despite the lack of observed morphological/ecological differences? One could argue that identifying and describing cryptic diversity can be important not only for more accurate biodiversity measures but also conservation management plans (i.e., evolutionary significant units; Ryder, 1986). For example, Fennessy et al. (2016) delimited four species of giraffe based on a genetic isolation criterion and placed special emphasis on conservation management of the northern giraffe Giraffa camelopardalis and its four recognized subspecies due to the severity of population declines when compared to other related species. In our case, the North lineage has been severely threatened by fires over the last 20 years compared with Central+South lineages. Specifically, more than half of the North lineage population occurrence records fall within a fire perimeter that occurred between 2000 and 2020, compared with approximately 20 percent for Central+South occurrence records falling within a fire perimeter. Failing to recognize the obvious genetic diversity in the North lineage could result in its loss due to lack of a management plan targeting their distribution in the Transverse Ranges or trying to manage all of A. icenoglei as one species could also result in not having adequate recognition and consequently protection for the North lineage.

Alternatively, one could argue that there is no practical value of recognizing genetically diverged lineages as separate species considering the lack of any visible diagnostic character/difference in ecological role (Freudenstein et al., 2016). Specifically, Freudenstein et al. (2016) argued that possessing both a unique ecological role and phenotypic differences are imperative when recognizing distinct species units. However, even this argument is rife with subjectivity; for example, how much phenotypic difference is enough to distinguish lineages as separate species? Also, it has been established that the speciation process is a continuum in which certain biological properties can be affected at different points along that continuum (Abbott et al., 2013; de Queiroz, 2007). Thus, it is feasible for geographically separated populations to accumulate enough genetic divergence for reproductive isolation despite still having morphological and ecological stasis. However, if one was to view species only in the context of a small snapshot in time (i.e., time-limited view of species; Freudenstein et al., 2016) and assumes reproductive isolation based only on genetic divergence, then that raises the question of what happens if/when secondary contact occurs with a sibling sister “species” or lineage (i.e., time-extended view of species; Freudenstein et al., 2016). One of the two options is possible if secondary contact occurs: (1) morphological and/or ecological differences could emerge to maintain reproductive isolation (reinforcement), or (2) hybridization occurs and genetic divergence between populations is eliminated via the effects of gene flow. Freudenstein et al. (2016) argue that viewing species as ecologically distinctive with historical gene flow combines both the temporal and phenotypic natures of species and alleviates the ambiguity of whether or not genetically diverged yet morphologically/ecologically homogenous lineages will remain diverged in the future. Thus, the most conservative taxonomic approach would be to require rejection of both genetic and ecological interchangeability for identifying separate cohesion species.

Studies spanning different animal taxa that have utilized CSC-based delimitation approaches have highlighted the importance of evidence for adaptive divergence when delimiting species (Bond & Stockman, 2008; Leaché et al., 2009; Newton et al., 2020; Rengifo-Correa et al., 2021). For mygalomorphs, Bond and Stockman (2008), the study upon which our CSC framework is based, delimited A. miwok and A. stephencolberti within the A. atomarius species complex based on mitochondrial data plus evidence of adaptive divergence (i.e., coastal dune habitats and lighter abdominal coloration). In a follow-up study, Garrison et al. (2020) found evidence of chemosensory-associated gene families under selection in dune endemics compared with their inland sister lineages, further elucidating patterns of ecological differentiation between coastal and inland sister species. Another example within mygalomorphs includes Newton et al. (2020) who initially identified five genetically distinct lineages within the Antrodiaetus unicolor species complex; however, genetic and ecological exchangeability assessments led to the delimitation of three species, not five, based on molecular, behavioral, and morphological data. In a similar study, Leaché et al. (2009) identified five phylogeographic groups within the coast horned lizard Phyrnosoma coronatum species complex based on molecular data, yet an assessment of climatic niche models and morphometrics of cranial horn shapes led to the delimitation of three species based on multiple operational criteria. Lastly, another example involves the difficult taxonomic status of kissing bugs within the Triatoma phyllosoma species group, where species limits have been hard to establish given occurrences of hybridization and cryptic diversity (Rengifo-Correa et al., 2021). Despite relatively low genetic divergence and the potential for hybridization, species within the T. phyllosoma complex can be distinguished based on morphological characters (i.e., head phenotype) and are all considered separate cohesion species.

Our analytical results separately inferred one to three species within A. icenoglei depending on the dataset and analysis used, but the final decision, arguably subjective, comes down to emphasizing mygalomorph life history characteristics and acknowledging limitations for each data type/analysis (discussed further below). The three species hypothesis was dismissed due to: (1) the less conservative niche equivalency test (Warren et al., 2008), (2) the possibility that the 50 percent locus completeness SNP dataset overly inflated cluster separation between Central + South, and (3) no obvious barrier to gene flow between Central and South lineages. The two species hypothesis is not substantiated based on morphological and ecological similarity between lineages, yet it is supported by rejecting genetic exchangeability as inferred by the VAE cluster separation with higher/more conservative locus completeness percentage datasets (75p and 90p) and a probable hard barrier to gene flow, the LA Basin, between North and Central+South. Although the LA Basin is likely impeding gene flow due to urban development and habitat fragmentation, the small likelihood of a potential corridor of habitat suitable for dispersal along the northern Basin rim/southeastern San Bernardino mountains cannot be completely dismissed (e.g., figure 1 in Vandergast et al., 2006). The one species hypothesis is supported by morphological and ecological data as well as an implementation of a supervised machine learning analysis on the 90p SNP dataset. Notably, the CLADES training model used in our study is potentially not appropriate for mygalomorphs, and the prevalence of morphological homogeneity (Bond & Stockman, 2008; Harvey et al., 2018; Hedin et al., 2013; Leavitt et al., 2015; Newton et al., 2020) and ecological similarity (Cooper et al., 2011; Hedin & Carlson, 2011; Rix et al., 2020) among mygalomorphs could obscure actual evolutionary diversity. The flowchart in Bond and Stockman (2008) suggests that rejecting genetic exchangeability for parapatric lineages, but not rejecting ecological interchangeability, can still potentially indicate separate cohesion species if niche conservatism is occurring. However, this view must be balanced with acknowledging that sparse, if any, evidence for adaptive divergence could indicate that reproductive isolation is not complete (i.e., ecological divergence is usually correlated with reproductive isolation; Freudenstein et al., 2016; Rissler & Apodaca, 2007), especially for parapatric lineages that still have the potential for gene flow in the future. Considering the lack of congruence across data types and analyses, we are taking the most conservative approach by retaining species boundaries within A. icenoglei until additional data types, both ecological and whole genomes, can be included for evaluating cohesion species identity.

4.3 Limitations of analyses and future prospects

We believe that the supervised machine learning analysis has limitations due to potential shortcomings with the training data set devised using unrelated taxa. Although we see great value in attempting to establish a training dataset integrating biologically/ecologically relevant characteristics, it is difficult to assess how applicable this dataset can be to other dispersal-limited taxa, especially across different taxonomic orders and biogeographical regions (Derkarabetian et al., 2022). First, the taxon Metanonychus, on which the training dataset was established, diverged approximately 25 mya, whereas the A. icenoglei sibling species complex likely diverged much later, which could artificially conflate deeper divergences with a predetermined “species cutoff” value, even if shallower species divergences are observed. Second, Metanonychus is found throughout the Pacific Northwest (Derkarabetian et al., 2019) whereas the A. icenoglei complex is found throughout the California Floristic Province, a biodiversity hotspot characterized by the intense complexity of geological, climatic, and topographic changes (Myers et al., 2000). One could argue that the overall complexity of the CAFP might influence the speciation process of low dispersal taxa in a different manner from how topographic changes in the Pacific Northwest would to the point that the genetic signatures may manifest differently. Specifically, as there are more topographical changes (both in number and intensity), the more chances there are for speciation through vicariance when compared to fewer/less drastic topography shifts (Badgley et al., 2017).

Our VAE analysis with the lower locus completeness dataset (50p) showed obvious separation between all three of the A. icenoglei lineages, whereas our higher locus completeness datasets (75p and 90p) retained only enough signal to maintain the North lineage as a separate cluster but not for Central or South lineages (Figure 4). VAE relies on the inherent structure present in the data (Derkarabetian et al., 2019), and previous studies have shown that VAE analyses have been heavily influenced by the filtering parameters for the SNP datasets (Martin et al., 2021; Newton et al., 2020). Specifically, if a lower threshold for locus completeness is allowed in a dataset the more likely it is to “over-split”, whereas more stringent filtering (i.e., a high threshold for locus completeness) can remove potentially important signal and “under-split” the amount of diversity. Because our higher locus completeness datasets retained the same clustering patterns, we are confident that they accurately reflect genetic divergence, and that the 50p dataset separation pattern for Central and South is an artifact of the filtering choice. Thus, it is important to be mindful of the potential filtering strategies for these SNP datasets, and best practices if utilizing VAE as a species delimitation method would be to use multiple filtering strategies to identify possible data artifacts versus actual structure.

There are known caveats for using niche-based distribution modeling approaches as a proxy for evaluating ecological interchangeability. First, it has to be acknowledged that large-scale ecological data, which are based on a very small time frame of 30 years (i.e., 1970–2000), used for building the SDMs potentially lacks the resolution needed for detecting very small-scale habitat differences which may be important for detecting adaptive divergence (Massatti & Knowles, 2014; Newton et al., 2020; Starrett et al., 2018). The microhabitat preferences for these spiders, which includes shaded ravines, north-facing slopes, and specific soil types (Bond, 2012), found within the heterogeneous landscapes throughout the CAFP are potentially not identified in the SDMs by even the best resolution available. Thus, our niche similarity tests using these models likely do not detect the potential for microhabitat niche divergence and consequently suggest the need for studies that assess fine-scale data on variables like temperature, precipitation, burrow features (e.g., size and depth), and soil composition.

Second, as discussed above, background region choice can heavily impact the results of niche similarity tests, thus incorporating multiple regions with biologically relevant constraints may provide a more rigorous application. Third, considering that our proxy of ecological interchangeability was only based on the abiotic factors contributing to niche space (i.e., bioclimatic variables and occurrence records to build an SDM), one could argue that there were other potential avenues of ecological divergence that could have been included in this study for a more robust evaluation of ecological interchangeability. There are potential biotic factors (e.g., competition with other taxa, difference in prey items across microhabitats, or non-overlapping breeding periods) that could distinguish lineages from one another. For example, other studies delimiting mygalomorph species have included behavioral traits when applicable (e.g., non-overlapping breeding periods; Hendrixson et al., 2015; Hendrixson & Bond, 2005; Prentice, 1997). Unfortunately, the lack of available natural history data for many fossorial mygalomorphs (Bond, 2012; Hedin et al., 2013; Starrett et al., 2017) have limited use of this type of data in species delimitation decisions.

Given these limitations, there are many potential avenues in which researchers can begin to bridge these gaps in knowledge. First, generating more datasets comprising low-dispersal taxa with varying divergence times and across other biogeographical regions that can be used to train models for supervised machine learning methods such as CLADES, will likely aid the robustness of this approach (Derkarabetian et al., 2022). Second, accumulating more natural history data for mygalomorphs will not only provide valuable general ecological information but may also be used as additional evidence in species delimitation. For example, pitfall trapping spiders in areas where occurrence records of each species/lineage of interest is well-known to collect specimens can be informative for both breeding period times and gut content analysis to identify prey items that are being ingested (i.e., can inform potential for ecological divergence). Third, the advent of assembled and annotated genomes for non-model taxa, specifically in Aptostichus, will likely pave the way toward utilizing these data not only for reconstructing evolutionary relationships but also identifying genes that contribute to potential adaptive divergence across the landscape (Johnson, 2018).

Overall, our study emphasized the efficacy of implementing a cohesion species-based delimitation approach across all taxa, but especially for assessing the potential of cryptic diversity. Using genome-wide UCEs in conjunction with morphological and ecological data to evaluate genetic and ecological exchangeability provided multiple independent lines of evidence that covered multiple biological properties potentially important in the speciation process. Specifically, this integrative approach underscored how different data types or approaches alone could either over- or under-split diversity estimates, yet taking them all into consideration led to a more robust species delimitation hypothesis within the A. icenoglei complex. Typically, such studies of taxa with extreme population structuring favor recognizing cryptic species, whereas herein, we have shown that an integrative approach, considering multiple lines of evidence, has the capacity to retain (lump) populations as a single species. Moreover, we reinforce the capability of the Cohesion Species Concept in providing both the conceptual and experimental framework for conducting such tests. Finally, our biogeographic analysis reveals that vicariance likely played a dominant role in the speciation process across the entire complex, further highlighting the impact of the complex geological, climatic, and topographical changes across the CAFP on speciation process.