Phylogenomics supports incongruence between ecological specialization and taxonomy in a charismatic clade of buck moths

2.1 Specimen sampling, DNA extraction and ddRAD

Specimen sampling of adult moths and caterpillars predominately followed that of Rubinoff et al. (2017). For most specimens, we conducted new DNA extractions to provide sufficient DNA for ddRAD library preparation, and only when additional material was not available did we use prior DNA extractions. We also included new specimens to expand sampling in several geographic regions. Species determinations followed Peigler and Stone (1989) and Lemaire (2002), and detailed specimen information is provided in Supporting Information Table S1. We extracted DNA from adult thoracic tissue or cross-sectional pieces of caterpillars and homogenized tissue in tissue lysis buffer with a 3.175-mm 18/10 stainless steel bearing using a FastPrep 24 homogenizer (MP Biomedical, Santa Ana, CA) at a speed of 4.0 m/s for 20 s. Tissue homogenate was then incubated with Proteinase K following manufacturer's recommendations (Macherey-Nagel, Düren, Germany) overnight at 55°C. We finished extractions using a KingFisher Flex-96 automated extraction instrument (Thermo Scientific, Waltham, MA) with NucleoMag tissue extraction kits (Macherey-Nagel; Düren, Germany) and the optional RNase A treatment, following manufacturer's recommendations, and eluted into 100 μl of Mag-Bind elution buffer. We quantified DNA by Quant-it Picogreen assay (Thermo Fisher; Waltham, MA) on a SpectraMax M2 plate reader (Molecular Devices; Sunnyvale, CA) and used a Gilson PIPETMAX 268 (Gilson; Middleton, WI) to normalize DNA to 4 ng/μl in 44.5 μl dH₂O.

We prepared ddRAD libraries following the method of Peterson et al. (2012) using ~178 ng of input DNA and restriction enzymes NlaIII and MluCI. The initial ligation was performed using 42 unique barcoded adapters, and a Blue Pippin electrophoresis unit (Sage Science; Beverly, MA) was used to size-select subpools of these 42 adapters with a 1.5% agarose gel cassette and “tight 450 bp” target size selection. We added Illumina i7 barcodes to each subpool in a final PCR and then performed a clean-up using a 1.5:1 ratio of polyethylene glycol containing solid-phase reversible immobilization beads/sample volume (DeAngelis, Wang, & Hawkins, 1995). We quantified these size-selected and cleaned products using a 2100 Bioanalyzer with the high-sensitivity DNA kit (Agilent, Santa Clara, CA) and generated a final library of 163 individuals pooled at equal molar ratios (including four blanks). This final library was sequenced using 100 bp single-end sequencing on an Illumina HiSeq 4000.

2.2 Data processing

We used the Stacks pipeline (Catchen, Amores, Hohenlohe, Cresko, & Postlethwait, 2011; Catchen, Hohenlohe, Bassham, Amores, & Cresko, 2013) to build catalog loci and call SNPs de novo. Unless specified, default settings and parameters were used for all data processing steps. First, we used process_radtags to demultiplex raw FASTQ files, discard reads with low-quality scores or uncalled bases and rescue reads that contained errors in the cut site or barcode. Then, we used denovo_map.pl to execute the main Stacks pipeline, requiring a minimum of three identical reads to create a stack and allowing for three mismatches between loci when building the catalog and when processing individuals (to accommodate higher variability between species in this data set). Finally, we used populations to generate two output formats from Stacks: vcf format containing individual SNPs for population genetic analyses and phylip format containing SNPs and flanking sequence (invariant sites of each catalog locus surrounding SNPs) for phylogenetic analyses. A minimum stack depth of eight reads per individual was required for both formats. For the vcf output, we used a population map assigning all individuals to a single population and then used vcftools version 0.1.14 (Danecek et al., 2011) for calculation of missing data per individual after excluding sites with >50% missing data. Based on this calculation of individual missingness, we then manually removed individuals with >80% missing data from the Stacks outputs and reran populations to create a new vcf file that was not biased by these poor-quality individuals and contained a single random SNP per catalog locus. We used vcftools again to create a final population genetic data file, excluding loci missing in >30% of individuals and those with a minor allele frequency (MAF) <1%. This MAF was chosen to accommodate several populations with low sample sizes (N < 6), where fixed unique alleles would have frequencies <5%. pgdspider version 2.1.0.3 (Lischer & Excoffier, 2012) was used to convert vcf format to other formats, and we refer to this data set as the “population genetic” data set. For the phylip output, we first excluded the same individuals as in the population genetic data set (identified with vcftools individual missingness calculation). We then used a population map assigning all individuals to unique populations and required loci to be present in >50% populations/individuals. We refer to this data set as the “phylogenetic” data set.

2.3 Population structure

We assessed population structure using multiple population genetic and phylogenetic approaches. First, we conducted Bayesian individual-based clustering of the population genetic data set using structure version 2.3.4 (Pritchard, Stephens, & Donnelly, 2000), which assigns individuals to genetic clusters (denoted by K) while maximizing linkage and Hardy–Weinberg equilibria. We ran 50 iterations of K = 1–20, each with 500,000 Markov chain Monte Carlo replicates (after 100,000 replicates of burn-in), and used the admixture model and correlated allele frequencies (Falush, Stephens, & Pritchard, 2003). We also conducted identical analyses using a location prior based on sampling locality, a hierarchical approach to assess substructure in the data (Vähä, Erkinaro, Niemela, & Primmer, 2007), and with the alternate ancestry prior and α = 1/10 (based on K = ~10 from preliminary analyses) as suggested by Wang (2016) for data sets with uneven sampling across populations. clumpak (Kopelman, Mayzel, Jakobsson, Rosenberg, & Mayrose, 2015) was used to average results across runs. We evaluated the optimal value of K using Ln Pr(X|K) (Pritchard et al., 2000), ΔK (Evanno, Regnaut, & Goudet, 2005) and the estimators introduced by Puechmaille (2016) (MedMedK, MedMeaK, MaxMedK and MaxMeaK, which we refer to as the “Puechmaille statistics” henceforth) implemented in structureselector (Li & Liu, 2017). For the latter, we used a population map corresponding to collection localities (generally, at the level of counties or states to avoid single individual populations) and a threshold of 0.5 (to accommodate high levels of admixed signatures found in preliminary analyses). We also used the population genetic data set to generate a NeighborNet phylogenetic network using splitstree version 4.14.6 (Huson & Bryant, 2006) with uncorrected distances.

We then used the phylogenetic data set to assess population structure using maximum-likelihood (ML) and Bayesian inference (BI) tree searches. For ML analysis, we used iq-tree version 1.6.0 (Nguyen, Schmidt, von Haeseler, & Minh, 2015), with 1,000 replicates of both ultra-fast bootstrap (Hoang, Chernomor, von Haeseler, Minh, & Vinh, 2018) and the Shimodaira–Hasegawa approximate likelihood ratio test (SH-aLRT; Guindon et al., 2010) to test node support, and used the best-fit substitution model predicted by ModelFinder (Kalyaanamoorthy, Minh, Wong, von Haeseler, & Jermiin, 2017). We used mrbayes version 3.2.6 (Huelsenbeck & Ronquist, 2001; Ronquist & Huelsenbeck, 2003) in parallel (Altekar, Dwarkadas, Huelsenbeck, & Ronquist, 2004) for BI analyses, with six runs of one million generations, each with four chains at default temperatures, sampling every 5,000 generations, and the substitution model as predicted for the ML analysis. We assessed convergence of the runs by monitoring the average standard deviation of split frequencies and the potential scale reduction factor (which should approach zero and one, respectively) and used tracer version 1.6 (Rambaut, Suchard, Xie, & Drummond, 2014) to assess effective sample sizes (which should be >200) after the analysis. We manually combined trees from independent runs, removed the first 25% of trees as burn-in and constructed a 50% majority rule consensus tree in paup version 4.0b10 (Swofford, 2002). figtree version 1.4.2 (Rambaut & Drummond, 2010) was used to visualize trees.

2.4 Descriptive statistics

We calculated general population genetic statistics (heterozygosity, inbreeding coefficients, etc.) with genodive version 2.0b27 (Meirmans & Van Tienderen, 2004), and private alleles per population with the poppr library (Kamvar, Brooks, & Grunwald, 2015; Kamvar, Tabima, & Grunwald, 2014) in r (R Core Team, 2018). We also used genodive to test for pairwise population differentiation using the molecular variance F_ST method (Excoffier, Smouse, & Quattro, 1992; Michalakis & Excoffier, 1996) with 10,000 permutations and to calculate pairwise Jost's D distances (Jost, 2008). Finally, we used genepop version 1.2 (Raymond & Rousset, 1995; Rousset, 2008) to calculate gene diversity among individuals in each population (1 − Q-inter). We applied Bonferroni corrections to tests with multiple comparisons.

3.1 Data properties

DNA was extracted from 137 individuals. However, initial digests produced inconsistent results for 23 individuals (confirmed by running digested products on a 1.5% agarose gel), so multiple DNA extractions and libraries were generated for these individuals, leading to a total of 163 individual libraries that were sequenced on a single lane of an Illumina HiSeq 4000. A total of 283.4 million reads were generated across all individuals, and 262.7 million of those were retained after removing reads of low-quality and ambiguous barcodes (mean per individual: 1.6 million reads). For individuals where we generated multiple DNA extractions and libraries, we selected the best library per individual by considering individual missingness calculated during the generation of the population genetic data set. This led to a total of 119 individuals being included in the population genetic and phylogenetic data sets. The population genetic data set contained 2,111 SNPs, and the phylogenetic data set contained 43,424 variable sites (11,881 Stacks catalog loci) across a 903,168-bp alignment.

3.2 Population structure

All analyses of population structure agreed on roughly the same pattern of differentiation between populations. At a broadscale, we observed high differentiation between three clusters corresponding to (a) Hemileuca artemis and Hemileuca nevadensis (b) Hemileuca maia (bog buck moth) from New York and (c) the remainder of the H. maia complex. However, at finer scales, we found support for additional clusters corresponding to the other described species and many geographically defined units (Figure 2). For the main structure analyses (all individuals, no location prior, default ancestry prior and α), ΔK supported K = 3, Ln Pr(X|K) plateaued at K = 11, and the Peuchmaille statistics generally supported K = 8 or 9. At the higher values of K, up to K = 11 where the finest-scale substructure is observed, the third aforementioned cluster is increasingly divided into distinct groups. At K = 11, Hemileuca lucina, Hemileuca peigleri and Hemileuca slosseri form distinct clusters and the remaining H. maia individuals cluster geographically: individuals from Florida, Iowa, Louisiana/Georgia, New York/Massachusetts and Wisconsin/Michigan/Nebraska form relatively distinct clusters (Figure 2). Some admixture was apparent in these groups, particularly the latter two groups, and the 11th cluster (represented by orange in the bar plot; Figure 2) was only represented at Q < 0.6. Above K = 11, no additional unique clades were observed, and the additional genetic clusters were represented by increased admixture in pre-existing clusters. structure analyses with a location prior and using the alternate ancestry prior and modified α resulted in virtually identical results as the main analysis, so we focus on the latter here. Hierarchical analysis of the main three clusters also resulted in virtually identical results as the K = 11 scenario; the only exception was the analysis of the H. artemis/H. nevadensis cluster, where K = 2 was supported, which separated the two species (Figure 2, inset). Given the overall pattern of hierarchical structure, results of other analyses (see below) and the current recommendations for structure analyses (see Janes et al., 2017 and references therein), we focus our discussion on the broadest-scale (K = 3) and finest-scale (K = 11) genetic differentiation that is supported in the data set (Figure 2). All structure results (including Ln Pr(X|K), ΔK, the Peuchmaille statistics and bar plots for each value of K) are provided in Supporting Information Figure S1.

The phylogenetic network analysis and ML/BI tree searches agreed with the population structure predicted by structure, for the most part (Figure 3). The highest differentiation (both in branch length and in distance measures) was observed for clusters of H. artemis/H. nevadensis and H. maia (bog buck moth) from NY. At finer scales, most of the same clusters observed in structure were highly supported or distinct, even some of the finest-scale differentiation such as between H. artemis and H. nevadensis (Figure 3). However, some small differences were found. Populations of H. maia from MA/NY and MI/NE/WI, which were mostly contiguous clusters in the structure analyses, were split in both the phylogeny and phylogenetic network (some MA individuals were distinct from the rest of MA/NY, and MI/WI clustered apart from NE). Hemileuca peigleri was paraphyletic with regard to H. slosseri; however, the clade containing these two species was strongly supported. Finally, in the phylogeny, we observed differentiation between H. maia from GA and LA; in the phylogenetic network, we found a similar pattern but several individuals fell outside of these main groupings (including one individual of H. peigleri, although this individual did have high missing data which would bias its distance measures) (Figure 3). Consensus trees for ML and BI were virtually identical, with the only main differences being one case of more ambiguous mid-range relationships with BI (see Supporting Information Figure S2); fully labelled trees are provided in Supporting Information Figure S2.

3.3 Descriptive statistics

We used results of all of the population structure analyses to create a synthetic hypothesis of fine-scale population divisions in the group and used these divisions for calculating descriptive statistics. These predominately correspond to the K = 11 structure results; however, some clusters were further split or joined based on the phylogenetic and network analyses, leading to a total of 12 population units (Figure 2). We named these units by species when applicable and otherwise using state abbreviations for divisions within H. maia (Table 1). Heterozygosity was relatively low for all populations, ranging from 0.009 to 0.093 (Table 1), although this is largely due to the relatively low MAF; increasing the MAF to 5% resulted in measures of heterozygosity that were roughly twice as high (Supporting Information Table S2). Both inbreeding and gene diversity were variable but relatively low, ranging from 0.029 to 0.640 and 0.021 to 0.112, respectively (Table 1). Pairwise F_ST values were very high (average = 0.363, Table 2); however, these values were likely inflated by overall low heterozygosity and a high number of private alleles across the data set: 766 of the 4,222 (18.1%) alleles were present only in a single population (Table 1). Jost's D values are independent from heterozygosity (Jost, 2008), and exhibited less inflated values (average = 0.049), but in some cases were impacted by low sample sizes in some comparisons, leading to values of zero (Table 2). Overall, F_ST and Jost's D exhibited similar patterns and mimicked the K = 3 results of structure, with the highest differentiation being found in comparisons of H. artemis/H. nevadensis and New York H. maia (bog buck moth) to the remainder of the H. maia populations (Table 2).

4.1 Drivers of speciation in the H. maia complex

While local adaptation and geographic isolation can both drive diversification, it is clear that their relative effects are remarkably heterogeneous across the members of this species complex. For instance, local adaptation to different host plant and habitat appears to be important in the isolation of Hemileuca lucina from the parapatrically distributed H. maia (Stamp & Bowers, 1986), akin to other well-known cases of selection-based ecological speciation across small geographic scales (e.g., Gasterosteus sticklebacks: Hatfield and Schluter (1999), Rhagoletis fruit flies: Filchak, Roethele, and Feder (2000) and Timema walking sticks: Nosil (2007)). The broad genetic differentiation of the H. artemis/nevadensis cluster in the West may also be explained by ecological diversification, as their exclusive feeding on Salix and Populus in riparian habitats is relatively unique. The eastern limits of this genetic entity and the extent of Salix/Populus-feeding in other regions are less clear (Scholtens & Wagner, 1995; Tietz, 1972) (discussed below).

The other instance of large genetic differentiation is less easily explained: Why does the bog buck moth exhibit such high relative genetic divergence compared to the other ecologically divergent populations? There are several potential explanations. First, the age of the lineage compared to the rest of the group could explain the higher divergence. However, bog buck moth is thought to be one of the more recent derivatives (post-glaciation) in the group (Rubinoff & Sperling, 2004; Tuskes et al., 1996), which is sensible given current understanding of the history of North American glaciation (e.g., Mickelson & Colgan, 2004); thus, greater age is unlikely to explain this divergence. Ecological divergence also fails to explain the high genetic differentiation, as there are populations in the western Great Lakes region with virtually identical morphology and ecology (e.g., Kruse, 1998), but no genetic diagnosability at the same hierarchical level. Therefore, the most likely explanation for high genetic differentiation in bog buck moth is its geographic isolation which limits gene flow. This explanation would also suggest that the other populations in the broadly distributed H. maia cluster are not physically separated to the same degree and are maintaining their ecological divergence in spite of gene flow. Taken together, our results provide a model for understanding both how local adaptation in an herbivore can arise and be maintained in the face of apparently strong gene flow and the importance of geographic isolation in generating broader genomic divergence, despite a lack of ecological divergence (Crispo, Bentzen, Reznick, Kinnison, & Hendry, 2006). While the varied nature of the diversification mechanisms shown in the H. maia species group has been seen, to a lesser extent, in other systems (e.g., Parchman et al., 2016), the complexity and variation in patterns of diversification are surprising and may be tied to evolutionarily important characteristics, such as mating systems, which are relatively indiscriminate in these moths (Collins, 2018; Peigler & Williams, 1984; Tuskes et al., 1996).

4.2 Species boundaries and taxonomy

The concept of what defines a species is fraught with inconsistencies and widely variable benchmarks (e.g., De Queiroz, 2007). Unfortunately, our results only further confirm the equivocal nature of such definitions. Specifically, the ecological divergence vs. genomic divergence contradictions suggest that there are serious inconsistencies in our concepts of species in the group, even when informed at the genomic level, to the extent that they may be counterproductive with respect to evolutionary concepts and applied uses (Drès & Mallet, 2002).

Taken at face value, the broadest scale of genetic differentiation could be used to suggest the elevation of bog buck moth in New York to species status, which would have significant management implications, since it is of conservation concern (see below). The same rationale would sink several described species, H. artemis, Hemileuca peigleri, Hemileuca slosseri and H. lucina (another species of conservation concern, see below), some of which show larger ecological and morphological divergence and distinctness than when comparing, for instance, the bog buck moth to other Menyanthes-feeding populations in the Midwestern Great Lakes region. Alternatively, at a finer genetic scale, one could argue that almost any of the small genetic clusters could be treated as distinct species or subspecies, based on varying combinations of characteristics tailored to suit each situation (e.g., genetic differentiation, host or habitat). Ultimately, even our genomics-informed view of population structure in this complex continues to highlight much of the subjectivity and debate present in the world of species/subspecies concepts (e.g., De Queiroz, 2007; Mallet, 2001; Mayden, 1997). Broadly speaking, there are obvious, important, ecologically functional differences between some populations in the H. maia group, but these differences are not broadly reflected in the genomes of the various populations and do not, apparently, inhibit ongoing gene flow between the groups. At what level such “genomic integrity” (Sperling, 2003) can be identified between genetic entities will be intrinsic to how these populations are recognized (Wu, 2001).

Despite the fact that we have sampled across the continental range of the group, there are still sampling gaps (particularly across the range of H. nevadensis s.l.: see below), and thus the potential for effects of unsampled populations on the analyses used here (“ghost populations” sensu Lawson, van Dorp, & Falush, 2018). Therefore, we are reticent to propose any of these changes formally. Rather, we hope that these results spur new interest in the relationships and population structure of the H. maia complex and geographically focused studies using genomic approaches such as that used here. Such focused treatments are particularly needed in the range of H. nevadensis. The eastern limits of this species, and its intergrade into H. maia, have remained elusive with regard to morphology, host plant and genetics (Henne & Diehl, 2002; Kruse, 1998; Legge et al., 1996; Rubinoff et al., 2017; Scholtens & Wagner, 1997; Tuskes et al., 1996). While we provide a fine-scale analysis of the species frontier in New Mexico and west Texas, where there is clear ecological and morphological distinction between H. artemis and H. slosseri, populations feeding on Salix spp. in Nebraska, which were always considered H. nevadensis based on morphology and habitat/host plant, are clearly H. maia (Figure 2). This phenomenon is repeated for populations in the Midwest where a priori identities were more ambiguous; it is clear that these populations are also H. maia. Comprehensive sampling throughout the Midwest and West will be needed to elucidate the taxonomic and evolutionary situation in this unsampled region.

4.3 Management implications

Two taxa in this group, H. lucina and the bog buck moth, are of conservation concern in the United States, and for the latter, Canada (Buckner et al., 2014; Environment Canada 2015; Legge et al., 1996; Rubinoff & Sperling, 2004), thus giving these results, and their interpretation, applied significance. While H. lucina is described as a distinct species and has clear ecological differences (host plant in particular), it is only genetically distinct at the finest scale of differentiation (Figures 2, 3). If we consider this level of genetic differentiation, then virtually every geographic cluster of populations in the species complex becomes equally distinct taxa. Thus, from a systematic perspective, we either disregard the ecological uniqueness of H. lucina or we recognize that every cluster of populations in the complex that we sampled represents a distinct entity, even those across the eastern third of the continent that exhibit very limited to no morphological and ecological differentiation. The bog buck moth complicates this decision because, while not yet described as a species, it is distinct at the broadest scale of genetic differentiation considered here, reflecting much deeper divergence than H. lucina. This deep divergence may be explained by the isolation of bog buck moth to the north of most H. maia populations, vs. the potential for genetic exchange between H. lucina and H. maia given their parapatric distribution throughout the range of H. lucina. As discussed above, this suggests that the genetic underpinnings for H. lucina's ecological specialization are maintained in the face of strong gene flow from H. maia, which has obscured the distinct nature of the former at all but the finest levels of genetic differentiation.

This discrepancy between taxonomic status and genetic divergence of H. lucina and bog buck moth presents a particular interesting implication for federal conservation protection, as invertebrates must be described subspecies or species to be federally protected in the United States (Waples, Nammack, Cochrane, & Hutchings, 2013). Thus, H. lucina would currently be eligible for federal protection, while bog buck moth would not, although state protections do not generally have the same limitation. Management of population units below the species level is generally accomplished through identification of evolutionarily significant units, which are populations exhibiting genetic differentiation and adaptive variability that would warrant protection (Crandall, Bininda-Emonds, Mace, & Wayne, 2000; Fraser & Bernatchez, 2001; Pearse, 2016; Shafer et al., 2015). The presented genomic results definitely satisfy the criteria of genetic differentiation for bog buck moth, and the use of Menyanthes trifoliata may satisfy that of adaptive variability. However, the latter is not exclusive to the genetic cluster identified as bog buck moth, as populations in Michigan (Scholtens & Wagner, 1997) and Wisconsin (Kruse, 1998) also feed on M. trifoliata. More comprehensive sampling of bog buck moth (particularly Canadian populations), and more populations in Michigan and Wisconsin, will be important to build on these results and our knowledge of the extent of broad- and fine-scale genetic differentiation between these populations.