Genomic and Phenotypic Delimitation of Species in a Temperate Aquatic Biodiversity Hotspot Authors: Daniel J MacGuigan, Adam Taylor, Ava Ghezelayagh, Julia E Wood, Jeffrey W Simmons, Jon M Mollish, Thomas J Near Published: 2026 ABSTRACT AbstractBiologists have relied on morphological characteristics to identify, define, and formally describe species for the past 250 years. The advent of phylogenetic species concepts and the introduction of molecular data have spawned new species delimitation methods applicable to a wide range of eukaryotic lineages. However, these approaches heavily emphasize genomic data, often overlooking phenotypic traits. We present and implement a species delimitation approach that utilizes genome-wide markers from ddRAD-seq and meristic morphological traits, which have long been used to identify and delineate fish species. Our methodology employs unsupervised machine learning to analyze morphological data without a priori species assignments, allowing phenotypic patterns to emerge independently from genomic-based species delimitation. We apply our combined genomic and phenotypic methodology to the freshwater systems of Southeastern North America, a biodiversity hotspot where conservation efforts are hampered by an incomplete knowledge of species diversity. Our investigation focuses on the darter clade Allohistium, a threatened lineage comprising two described species. Through phylogenomic, population genetic, and phenotypic model comparisons, we provide evidence supporting the delimitation of a third species of Allohistium, which we formally describe. Our approach shows how unsupervised machine learning can reveal cryptic morphological diversity that might otherwise be obscured by taxonomic preconceptions. This study demonstrates that model testing using diverse lines of evidence yields a more comprehensive, data-driven hypothesis of species diversity. Identifying independent evolutionary lineages that merit recognition as distinct species is one of the fundamental goals of systematic biology. This process has long been hampered by disagreement about species concepts and species delimitation methodologies (De Queiroz 2007). There is broad consensus that interrogating multiple lines of evidence results in a more robust assessment of species-level biodiversity (Carstens et al. 2013). However, these lines of evidence are only as good as the tools available to analyze them. The use of genetic data is widespread in species delimitation studies (Near et al. 2018; Parker et al. 2022; Kim et al. 2023; Wood et al. 2023; Vences et al. 2024). Genetic data can be collected and analyzed in largely the same manner regardless of the organism. Additionally, genomic data sets provide substantial statistical power to infer evolutionary history. These attractive features have spurred the development of many methods for species delimitation using genetic data. Genetic species delimitation methods typically focus on “objective” assignment of individuals to lineages and statistical comparison of alterative delimitation models (Grummer et al. 2014; Leaché et al. 2014, 2019; Yang 2015). The performance of these genetic species delimitation methods can be confounded by many evolutionary processes such as isolation-by-distance (Burbrink and Ruane 2021; Hillis et al. 2021; MacGuigan et al. 2022), incomplete lineage sorting (Thomé et al. 2012; Weber et al. 2019), and introgressive hybridization (Jackson et al. 2017; Quattrini et al. 2019; Kim et al. 2022). Despite the rapid growth of DNA sequencing for non-model organisms, phenotypic information is still the primary type of data used to discover and delimit species in the early 21st century (Cadena and Zapata 2021). The use of phenotypic data to identify species of fishes relies on an operational species concept that originates in the early 19th century. Investigators attempt to identify specific phenotypic traits that exhibit divergent measures of central tendency between species, such as differences in mode, median, or mean trait values. The goal is to identify traits that could be used to diagnose individual specimens to species. More recently, multivariate approaches have become a common supplemental method used alongside the examination of individual trait distributions (Sokal and Sneath 1963). Typically, multivariate morphospaces are examined for the (often qualitative) degree of overlap between species (Zapata and Jiménez 2012). Both approaches rely on subjective, a priori assignment of individuals to species and lack any model comparison framework. Simply identifying traits with different central tendencies or graphically examining multivariate morphospaces is not sufficient for accurate species delimitation (Cadena et al. 2018). There is a clear need for objective, phenotypic approaches that avoid pitfalls of a priori lineage assignment and compare the fit of different species delimitation models. However, developing generalizable species delimitation tools for phenotypic data is challenging, as taxonomists focus on different types of morphological traits for different clades. The past few decades have seen the development of several promising phenotypic species delimitation methods that are more robust and thorough than traditional taxonomic approaches. These include threshold-based methods (e.g., Wiens and Servedio 2000; Ezard et al. 2010; Tobias et al. 2010), Bayesian phylogenetic approaches (Solís-Lemus et al. 2014), and normal mixture models (Zapata and Jiménez 2012; Edwards and Knowles 2014; Cadena et al. 2018). Yet, not unlike traditional taxonomic approaches, most of these methods require a priori assignment of individuals to species and/or do not allow comparison of alternative delimitation models. Additionally, with one major exception (Solís-Lemus et al. 2014), these methods are not readily available as software packages, limiting their widespread adoption by practicing systematists. More recently, machine learning approaches have shown promise for species delimitation without the need for a priori taxonomic assignment (Karbstein et al. 2024), but most published methods only utilize genetic data (Pei et al. 2018; Derkarabetian et al. 2019, 2022; Smith and Carstens 2020; Martin et al. 2021, but see Yang et al. 2022; Pyron 2023). Here, we propose a new method for species delimitation using multivariate morphological data sets. We use unsupervised machine learning to identify phenotypic clusters and a matrix similarity algorithm (Jakobsson and Rosenberg 2007) to compare the fit of the clusters to the underlying delimitation model. This approach is straightforward, requiring minimal computational resources, and is applicable to any multivariate morphological data set for which principal component analysis (PCA) is appropriate. We have released our method, PhenoDelimit, as a user-friendly R package on GitHub (https://github.com/dmacguigan/PhenoDelimit) with fully reproducible example code. In this paper, we deploy our new phenotypic delimitation method on the Allohistium darter species complex. Darters (Percidae: Etheostomatinae) are a clade of more than 230 described species of freshwater fishes endemic to eastern North America (Page et al. 2023). Although the genus Allohistium diverged from its closest relatives (the genus Nothonotus) approximately 24 million years ago (Ma) (MacGuigan and Near 2019), it contains only two described species (Powers et al. 2012). Allohistium is phenotypically divergent from other darters and experienced ancient introgression approximately 20–24 Ma (Page 1981; Kuehne and Barbour 1983; Shepard and Burr 1984; MacGuigan and Near 2019). The two recognized species of Allohistium are geographically isolated in the Tennessee and Cumberland river basins (Fig. 1), both hotspots of aquatic biodiversity (Warren and Burr 1994; Warren et al. 2000; Elkins et al. 2019). Phylogenetic analyses of mitochondrial DNA (mtDNA) suggest that A. cinereum is paraphyletic, with populations from the Duck River system resolved as the sister lineage of a clade containing A. maydeni in the Cumberland River system and populations of A. cinereum from the Elk River and upper Tennessee River system (Powers et al. 2004, 2012). Additionally, there is morphological variation among A. maydeni in the Cumberland River system, populations of A. cinereum in the upper Tennessee and Elk River systems, and populations in the Duck River system (Shepard and Burr 1984; Powers et al. 2012). Paraphyly and morphological variation suggest the presence of an additional undescribed species of Allohistium. All species and populations of Allohistium are listed as either vulnerable or endangered (Jelks et al. 2008 ), imperiled as a consequence of habitat loss driven by agriculture and urbanization (Shepard and Burr 1984; Etnier and Starnes 1993; Boschung and Mayden 2004; Powers et al. 2004; Compton and Taylor 2013). The development of effective conservation policies and practices for Allohistum is dependent on an accurate delimitation of the species-level biodiversity in the clade. Allohistium sampling map. Diamonds indicate localities in the meristic data set and circles indicate localities in the genetic data set. Stars indicate type localities for each species. Localities where Allohistium is likely extirpated are circled in black. Specimen photo is Allohistium cinereum (YFTC 38863, YPM ICH 034422). Scale bar in km is for the map, scale bar in cm is for the Allohistium specimen. We assembled a genomic-scale DNA sequence data set for species and populations of Allohistium using double-digest restriction enzyme DNA sequencing (ddRAD-seq). Additionally, we collected data for 11 meristic traits and 5 squamation (scale pattern) characters from 259 individuals, sampling all known extant populations of Allohistium (Fig. 1). These are phenotypic characters traditionally used to discover and delimit fish species (Page 1983; Hubbs et al. 2004). Phylogenetic, population genetic, and multispecies coalescent analyses identify three deeply divergent lineages of Allohistium delimited here as species that exhibit no evidence of gene flow. Our new phenotypic species delimitation method identifies the same three species as the best fit of all tested models. Analysis of morphological data sets for 13 other darter sister species pairs shows that the model fit of the three Allohistium species is comparable to other described species. Based on these results, we formally describe a new species from the Duck River basin: Allohistium anas. Our results demonstrate that PhenoDelimit is a new and useful tool that allows objective species discovery through exploratory clustering and species delimitation through model comparison using morphological data sets. Materials and Methods Collection, Sequencing, and Assembly We collected tissue samples and voucher specimens from 42 unique localities encompassing the range of Allohistium cinereum and A. maydeni (Fig. 1). Our genetic data set included 97 specimens of Allohistium (Supplementary Table S1). In addition, we included one sample from two outgroup species, Etheostoma mariae and E. fusiforme (MacGuigan and Near 2019). Etheostoma fusiforme was sequenced with our Allohistium individuals, whereas previously published E. mariae ddRAD-seq reads were downloaded from the National Center for Biotechnology Information (NCBI) Short Read Archive (SRR21949587, MacGuigan, Krabbenhoft, et al. 2023). With all appropriate scientific collection permits and following IACUC protocol number 2018-10681, individuals were euthanized in MS-222 and tissue samples were taken, usually from the right pectoral fin. Tissues were preserved in 95% ethanol at −20 °C. Voucher specimens were preserved in buffered 10% formalin, with long-term storage in 70% EtOH at the research collection of fishes at the Yale Peabody Museum (YPM). We extracted DNA using a standard Qiagen DNEasy Blood and Tissue Kit (Qiagen, Hilden, Germany) according to the manufacturer's instructions. DNA extractions were cleaned prior to library preparation using ethanol precipitation: 3M sodium acetate (pH = 5.2) was added equal to 10% of the volume of the DNA extraction, followed by 100% ethanol equal to 2.5 times the total volume of DNA. The extractions were incubated for 10 min at −80 °C, then centrifuged for 30 min at 8000 relative centrifugal force (RCF). The supernatant was discarded and the DNA pellet was washed with 250 μL of cold 70% ethanol. Samples were centrifuged again for 5 min at 8000 RCF and the supernatant was discarded. The pellet was air dried for approximately 15 min and resuspended with DNAse-free water. We performed a modified version of the ddRAD-seq protocol outlined in Peterson et al. (2012) and Poland et al. (2012). A total of 200 ng of DNA was digested from each sample for 8 h at 37 °C using the restriction enzymes PstI and MspI. We visually checked samples for complete digestion using a 2.5% agarose gel. The samples in this study were divided between two 96-well plates. Custom adapters were added using T4 DNA ligase and an incubation period of 3 h at 22 °C. The adapters contained a set of 96 unique 8–10 bp barcodes with at least two mutational differences between any pair of barcodes to minimize issues when demultiplexing. For each plate, all samples were pooled and cleaned using a QIAquick PCR purification kit (Qiagen, Hilden, Germany). Twelve rounds of PCR were performed, each cycle with 30 s at 98 °C, 30 s at 62 °C, and 30 s at 72 °C. Following the 12 cycles, the samples were held for 10 min at 17 °C. After PCR, sets of 96 samples with unique barcodes were pooled to perform a 300–500 base pair size selection using a Blue Pippen 2% agarose gel cassette (Sage Science, Beverly, MA, USA). We confirmed the efficacy of size selection using an ABI Bioanalyzer High Sensitivity DNA assay (Agilent Technologies, Santa Clara, CA, USA). Size selection and fragment analysis were performed at the Yale DNA Analysis Facility on Science Hill. We sequenced each 96-sample genomic library using one lane of Illumina HiSeq 4000 with 100 base pair single-end reads at the University of Oregon Genomics and Cell Characterization Core Facility. Raw read quality was inspected using FastQC (Andrews 2010). We performed reference-based assembly of the ddRAD-seq data using iPyrad v.0.9.85 (Eaton and Overcast 2020). There are currently no genomes available for Allohistium; therefore, we used the chromosome-level E. perlongum genome assembly (MacGuigan et al. 2023). Illumina reads were demultiplexed by iPyrad step 1, then trimmed using CutAdapt (Martin 2011). We trimmed reads to 100 bp and applied three rounds of trimming to remove adapter contamination and restriction cut sites (“-b TGCAG”, “-a TGCAG”, and “-a CCGAGATCGGAAGAGC”). Trimmed reads were passed back to iPyrad for the remaining assembly steps. For each locus, we required a minimum average depth of 6x, a maximum average depth of 10,000x, and a maximum of 5% heterozygous or ambiguous base calls. For each single-locus alignment, we allowed a maximum of eight indels and a maximum of 50% shared polymorphic sites. Additional assembly parameters are described in Supplementary Table S2. Exploratory Phylogenetic and Population Genetic Analyses We performed concatenated phylogenetic analyses of the Allohistium ddRAD-seq data set using IQTree v.2.1.1 (Minh et al 2020). To assess the impact of missing data on phylogenetic inference, we used iPyrad to assemble three data sets with different minimum samples-per-locus thresholds: 80 of 99 individuals (<20% missing data), 90 of 99 individuals (<10% missing data), and 95 of 99 individuals (<5% missing data). We used a general time reversible (GTR) + gamma nucleotide substitution model for all ddRAD-seq analyses. We ran each tree search until 100 unsuccessful tree search iterations were completed. To assess topological support, we also performed 1000 ultrafast bootstrap replicates (Hoang et al. 2018). We assembled several sets of single nucleotide polymorphisms (SNPs) for population genetic analyses. Using VCFTools (Danecek et al. 2011), we retained only biallelic SNPs (“--max-alleles 2”) and removed sites that contained more than 5% missing data (“--max-missing 0.95”). We also removed singletons (“--singletons”) and non-polymorphic sites (“--mac 1”) because these markers can confound inference of population structure (Linck and Battey 2019). We thinned our data set to include a single SNP per 10,000 base pair window (“--thin 10000”) to minimize the effects of physical linkage on our inference of population structure. Although the above filtering removed SNPs with large amounts of missing data, some individuals in our data set still contained large proportions of missing data. Therefore, we used the impute function in the R package LEA to replace missing genotypes with a random value weighted by the observed genotype probabilities (Frichot et al. 2014). We compared population structure results with imputed versus unimputed data. To examine population structure, we utilized the snmf function in the R package LEA v.3.4.0 (Frichot and François 2015). We performed 10 repetitions for each ancestral population (K) value from K = 1 to K = 20. We set the regularization parameter (alpha) to 10, the tolerance to 0.00001, the maximum number of algorithm iterations to 1000, and the percentage of masked genotypes for cross-entropy validation of K values to 0.05. For each value of K, the replicate with the minimum cross-entropy score was selected for further visualization and comparison with other clustering methods. A cross-validation approach using cross-entropy criteria was performed to determine the optimal value of K. In addition to ancestry coefficients, we reconstructed an effective migration surface for Allohistium using FEEMS v.1.0.0 (Marcus et al. 2021). We converted the VCF file used for LEA to bed format using PLINK v.1.90 (Chang et al. 2015). We generated a triangular discrete grid (encompassing Tennessee and Kentucky, USA) using the R package dggridR v.3.0.0 with a resolution of 11 and an ISEA projection (Barnes et al. 2022). To account for the dendritic structure of river basins, we used QGIS v.3.22 to crop the grid at Hydrologic Unit Code (HUC) 6 watershed boundaries (https://www.usgs.gov/national-hydrography/watershed-boundary-dataset). We then manually reconnected the grid where rivers cross watershed boundaries. This custom grid prevents direct migration between HUC-6 watersheds except by following the river course. Lastly, this gird was trimmed to the maximum extent of the Cumberland River and Tennessee River watersheds prior to running FEEMS. To identify the optimal smoothing parameter (lambda), we performed cross-validation using the FEEMS run_cv function with a set of 30 lambda values evenly spaced on a log scale from 1e−6 to 1e3. The lambda value with the smallest cross-validation L2 error was selected for the final FEEMS analysis. To estimate divergence times, population sizes, and migration rates, we performed a demographic analysis using GADMA v.2.0.4 (Noskova et al. 2020) with the moments v.1.3.1 inference engine (Jouganous et al. 2017). The input model was based on our phylogenetic and population structure results, dividing Allohistium into five populations. The model allowed asymmetric migration between all five extant population pairs. To address missing data and achieve reasonable runtimes, we used the site frequency spectrum (SFS) down-projection method of Marth et al. (2004) as implemented in easySFS (https://github.com/isaacovercast/easySFS), projecting down to 10 individuals per population. The SFS was generated from a thinned (one SNP per 10,000 base pair window) data set with a maximum of 20% missing samples per ddRAD locus. The effective sequence length (L) used to construct the SFS was specified as 391,220 bp based on the formula (X–Y)/X * Nseq, where X is the number of variable sites (171,832), Y is the number of SNPs removed by filtering (155,009), and Nseq is the total length of all sampled ddRAD loci (3,995,963 bp). For our demographic analyses, we assumed a generation time of 2 years (Shepard and Burr 1984) and a mutation rate of 2.26e−09 mutations per site per generation (based on germline mutation rates for a closely related perciform fish, Larimichthys crocea, Bergeron et al. 2023). For global parameter optimization, we used the “SMAC_BO_combination” algorithm with 100 independent replicates and a maximum of 200 evaluations per replicate. DemesDraw was used to visualize the resulting best-fit demographic model (Gower et al. 2022). Genetic Species Delimitation To explicitly test alternate species delimitation models, we calculated the genealogical divergence index (GDI) for all possible Allohistium lineages that might constitute species. GDI is heuristic metric bounded between 0 and 1 that is a function of branch length (τ) and genetic diversity (θ) (Jackson et al. 2017; Leaché et al. 2019). GDI values were calculated using equation 7 of Leaché et al. (2019). We used iPyrad to assemble a ddRAD-seq data set consisting of five individuals from each of the five major Allohistium lineages, as defined by the K = 5 population structure (see Results). We retained only loci that contained data for all 25 individuals with a minimum length of 80 base pairs and a maximum of 10% missing data. This resulted in a data set of 3924 loci. To estimate θ and τ, we used the A00 algorithm (fixed guide tree) of BPP v.4.1.4 (Rannala and Yang 2017). We used the R package PopGenome v.2.7.5 (Pfeifer et al. 2014) to estimate average nucleotide diversity within populations (Nei 1987) as a basis for the θ prior [IG(3, 0.0006913524)]. Additionally, we used the dist.dna function in the R package ape v.5.7-1 (Paradis and Schliep 2019) to calculate the uncorrected pairwise genetic distance between two specimens (AcinCR_ELK_TN and AmayBP_OBEY_TN), which capture the crown node of Allohistium. This genetic distance was used to parameterize the BPP τ prior [IG(5, 0.0137251)]. Importantly, none of the samples used to parameterize the priors were included in the BPP analysis. We performed four sets of analyses using the maximum likelihood phylogeny as the guide tree. These analyses split Allohistium into maximally five species and minimally two species. For each set of analyses, we performed four replicate runs with 500,000 generations of burn-in followed by 1,000,000 Markov chain Monte Carlo (MCMC) generations, sampling every 10 generations. MCMC traces were visually inspected for convergence and sufficient mixing (effective sample size [ESS] > 200). Posterior distributions of τ and θ were summarized for each lineage across all replicate runs. The pipeline used to perform BPP analyses and calculate GDI is available in a public GitHub repository (https://github.com/dmacguigan/gdiPipeline). Morphological Data Collection We collected meristic trait data from 259 specimens of Allohistium: 72 specimens of A. cinereum from the upper Tennessee River system, 18 specimens of A. cinereum from the Elk River in the lower Tennessee River system, 109 specimens of A. maydeni from the Cumberland River system, and 60 specimens of A. cf. cinereum from the Duck River system (Supplementary Table S3). The sampling locations of specimens examined for morphological comparisons are shown in Figure 1 and details on sampling locations are given in Supplementary Materials. The numbers of scale rows and fin elements were determined from each specimen as outlined in Hubbs et al. (2004)and Page (1981), with the exception of the number of transverse scale rows, which was counted as described by Page (1983, p. 16, Fig. 2). Morphological Species Delimitation We developed an unsupervised machine learning approach for morphological species delimitation. In this study, we apply the method only to meristic trait data. However, the methodology described below is applicable to other types of phenotypic data, such as landmark morphometrics. The data processing pipeline is available as an R package, PhenoDelimit (https://github.com/dmacguigan/PhenoDelimit). To compare the fit of different delimitation models, we first perform discriminant analyses of principle components (DAPC), which was originally developed to assess population genetic structure (Jombart et al. 2010). The DAPC approach is analogous to the method STRUCTURE (Pritchard et al. 2000) in that it generates individual assignment probabilities to different genetic clusters. The authors of DAPC note that their method could be applied to any quantitative data (Jombart et al. 2010). Here, we develop a pipeline for applying DAPC to morphological data sets. The DAPC method requires a priori assignment of individuals to groups. In PhenoDelimit, users can specify group assignments based on the underlying delimitation model with dapc_clumpp(apriori=TRUE). Alternatively, users can test how well the underlying structure of the morphological traits fit their delimitation model by assigning individuals using k-means clustering (dapc_clumpp(apriori=FALSE)). The k-means clustering method is an unsupervised machine learning algorithm that minimizes within-cluster variance and maximizes between-cluster variance. This allows the generation of group assignments without using prior knowledge based on genetics, geography, etc. By setting the number of clusters equal to the number of species, we can then compare how well unguided clustering of morphological data fits the a priori species delimitation model. We use the CLUMPP FullSearch algorithm (Jakobsson and Rosenberg 2007) to compare the fit of different species delimitation models. CLUMMP was originally designed for post-processing of population structure analyses where there can be multiple distinct solutions or arbitrary cluster label switching among replicate runs. The CLUMPP algorithm maximizes the similarity between assignment probability matrices using an average pairwise similarity score, H’, which is scaled between 0 and 1. An H’ value of 1 indicates a perfect match between the matrices. PhenoDelimit calculates H’ by comparing a matrix of DAPC assignment probabilities based on unguided k-means clustering versus a matrix where group membership coefficients perfectly match the delimitation model. If species in a delimitation model have strongly divergent trait distributions, then we expect these assignment probability matrices to be similar, and CLUMPP will return a high H’. We can compare H’ values for different species delimitation models to determine which model most closely fits the structure of the underlying data. The model with the highest H’ is considered the best-fit species delimitation. However, if all species delimitation models poorly fit the observed morphological data, we would not want to simply select the model with the highest H’. Therefore, we implemented a statistical permutation test in PhenoDelimit. For each species delimitation model, the morphological data are permuted a set number of times, then run through the same DAPC with k-means clustering procedure. We use CLUMPP to calculate H’ for each of the permuted data sets, then compare the observed H’ to the permuted distribution. If the observed H’ is not significantly larger than the permuted distribution, that indicates the morphological data have little structure corresponding to the proposed species. These delimitation models are rejected. Once we identify the best-fit species delimitation model, individual assignment probabilities can be visualized as a bar plot, similar to the ubiquitous “STRUCTURE plots” used in population genetics. This visualization helps identify individuals or subpopulations exhibiting morphological intermediacy. Additionally, we can examine which morphological traits contribute most to each discriminant function. By comparing the magnitude and direction of variable loadings to a scatter plot of the discriminant functions, we can identify potentially diagnosable traits for different species in our model. We applied the PhenoDelimit pipeline to a data set of 11 meristic traits and 5 squamation characters sampled from specimens of Allohistium. We tested six delimitation models. Model 1 matches the current two-species taxonomy, with A. maydeni (Cumberland and Big South Fork) versus A. cinereum (Duck, Elk, and upper Tennessee). Although model 1 includes a paraphyletic species, models 2 through 6 include only putative species that are monophyletic. Model 2 matches the two deepest clades in our phylogeny (Duck, Cumberland, and Big South Fork versus Elk and upper Tennessee). Model 3 is a three-species scenario that splits populations in the Duck River system (A. cf. cinereum) from populations of A. cinereum from the Elk River and upper Tennessee River system. Model 4 is a four-species version of Model 3 that divides the populations in the Elk River and upper Tennessee River system into separate species. Model 5 is also a four-species version of model 3 that splits the populations of A. maydeni from Big South Fork and remaining Cumberland River tributaries into separate species. Model 6 is a five-species scenario combining model 4 and model 5. In addition to the Allohistium data set, we ran PhenoDelimit on a set of meristic data sets for 13 sister species pairs that vary in their degree of genetic and morphological divergence (Ghezelayagh et al. 2025). For these sister species pairs, we did not perform model testing. Instead, we analyzed the two-species model for each species pair. The only exception was Nothonotus microlepidus and N. sanguifluus, which were grouped as a single lineage in comparison with N. starnesi. All traits were scaled and centered for the initial PCA. No missing data were allowed in our data sets. We retained all principal component (PC) axes for k-means clustering and performed 1,000,000 iterations of the algorithm for 1000 random starting centroids. We then performed discriminant analysis three times for each delimitation model, retaining the first set PC axes that explained 70%, 80%, or 90% of the cumulative variance. For each model, we performed a set of 1000 permutations, retaining PCs explaining 80% of the cumulative variance for the discriminant analyses. Results Taxonomy Allohistium anas Near and MacGuigan new species Cinder Darter Zoobank registration: urn:lsid:zoobank.org:act:7114E668-F9FD-4B47-8EA8-8B2C54528CBA Holotype. YPM ICH 038260 (Fig. 2), adult male, 85.7 mm standard length (SL). Duck River at river mile 221.3, approximately 200 m downstream of US Highway 231 bridge crossing, latitude 35.48362, longitude −86.46481, Bedford County, Tennessee, USA, 8 April 2020. Live photographs of Allohistium specimens. Photo credit to Jon M. Mollish. YPM = Yale Peabody Museum, YFTC = Yale Fish Tissue Collection. Paratopotypes. YPM ICH 034409, 3 females 71.4–80.2 mm SL and 3 males 53.4–78.4 mm SL collected with the holotype. Paratypes. YPM ICH 032927 (Fig. 2), 6 females 74.5–91.0 mm SL. Garrison Fork at Cannon Road bridge crossing, latitude 35.4934, longitude −86.3401, Bedford County, Tennessee, USA, 2 August 2018. YPM ICH 032852, 3 females 75.0-80.5 mm SL and 2 males 87.5-89.0 mm SL. Garrison Fork at Haley Road Crossing, latitude 35.486325, longitude -86.335796, Bedford County, Tennessee, USA. UT 91.4127, 1 female 63.4 mm SL and 2 males 36.4–52.9 mm SL. Duck River approximately 100 m upstream of Lillard’s Mill, latitude 35.58668, longitude −86.78471, Marshall County, Tennessee, USA, 30 May 1990. INHS 79389, 1 female 57.0 mm SL. Buffalo River at mouth of Grinders Creek, latitude 35.4628, longitude −87.53555, Lewis County, Tennessee, USA, 14 April 1978. Diagnosis. Allohistium anas differs from Allohistium cinereum and Allohistium maydeni by a higher number of lateral line scales (Supplementary Table S4), modally 43 versus 42 vertebrae (Shepard and Burr 1984), a larger proportion of individuals with greater than 80% of the cheek covered with scales, and more than 50% of the individuals with 10–40% of the nape covered with scales. In addition, Allohistium anas is never resolved as the sister lineage of Allohistium cinereum sensu stricto in mtDNA gene trees (Powers et al. 2004, 2012) or phylogenomic analyses of ddRAD-seq loci (see below). Allohistium maydeni differs from Allohistium anas and Allohistium cinereum in having modally 11 versus 12 dorsal fin spines (Supplementary Table S5; Powers et al. 2012), 12 versus 13 dorsal fin rays (Supplementary Table S6; Powers et al. 2012), and the presence of conspicuous red coloration on the skin covering the upper and lower oral jaws (Fig. 2). Etymology. The species name anas is from the Latin word for duck, in reference to the distribution of the species in the Duck River system. The common name Cinder Darter is in reference to the common name of A. cinerum, the Ashy Darter. Distribution and habitat. Allohistium anas is endemic to the Duck River system and is distributed in the mainstem of the Duck River from approximately the confluence with Garrison Fork downstream to Columbia, Tennessee, the lowermost portion of Garrison Fork, the Buffalo River from the confluence with Grinders Creek, Lewis County, Tennessee downstream to the confluence with Mill Creek, Wayne County, Tennessee. As reported for other species of Allohistium (Shepard and Burr 1984; Etnier and Starnes 1993; Compton and Taylor 2013), we encountered individuals of Allohistium anas in pools ranging in depth from 0.5 to 2 m with slow current, bedrock and gravel substrates mostly free of silt, and associated with large boulders and cobble. Specifically, A. anas is associated with rubble piles positioned horizontally near banks in areas of low flow. With the exception of the Buffalo River, it is distributed within the Nashville Basin, which is characterized by outcrops of Ordovician-age limestone, which are often in the form of rubble piles alongside river banks. The Buffalo River is located within the Western Highland Rim where streams are characterized primarily by chert gravel and sand substrates. Areas of underlying exposed limestone provide rubble piles used by A. anas in the Buffalo River. ddRAD-seq Data Set We sequenced 97 Allohistium specimens from 42 unique localities (Fig. 1; Supplementary Table S1), with a median of two specimens per locality (minimum = 1, maximum = 9). All individuals in the ddRAD-seq data set contained at least 618,000 reads (mean = 3.6e6 reads, SD = 1.6e6 reads) with an average of 24.9X coverage per locus (SD = 145.9X). iPyrad assembled a total of 368,284 ddRAD-seq loci, which were used to construct all phylogenetic and population genetic data sets. See Supplementary Table S1 for detailed sequencing statistics for each individual. Phylogenetics and Phylogeography To reconstruct the evolutionary history of Allohistium and compare the effects of missing data on tree inference, we used three ddRAD-seq data sets with different minimum samples-per-locus thresholds: 80 of 99 individuals (42,890 loci, 3,995,963 bp, 18.46% missing sites), 90 of 99 individuals (22,673 loci, 2,212,213 bp, 14.56% missing sites), and 95 of 99 individuals (11,241 loci, 1,118,698 bp, 12.14% missing sites). Analyses of all three data sets resulted in identical phylogenies at deeper nodes, with some differences at the shallowest intraspecific nodes (Supplementary Fig. S1). Therefore, we hereafter refer only to the phylogenetic results from analysis of the largest ddRAD-seq data set (42,890 loci) unless otherwise noted. The maximum likelihood tree for Allohistium was well-resolved with moderate to high bootstrap support values for most nodes (mean BSS = 76%, median = 88%). Our analyses resolve three main clades: A. cinereum from the Elk River and upper Tennessee River basin, A. anas from the Duck River basin, and A. maydeni from the Cumberland River system (Fig. 3A, B). Each of these clades is resolved with 100% bootstrap support. Notably, as currently delimited, A. cinereum is paraphyletic. Population structure and phylogeny of the Allohistium species complex. (A) Map shows pie charts of ancestry coefficients averaged across individuals for each locality. (B) Phylogeny is the maximum likelihood tree from an analysis of 42,890 ddRAD loci. Nodes with 95% or higher bootstrap support are unlabeled, and nodes with less than 70% bootstrap support are collapsed. Branch lengths are in units of expected substitutions/site. Bar plots show individual ancestry coefficients for between three and five ancestral populations (K). Order of individuals in the bar plot matches the order of those in the phylogeny. (C) Effective migration surface for the Allohistium species complex. Black lines represent HUC 6 watershed boundaries, which are treated as hard barriers to migration. Blue edges indicate higher than average effective migration, and orange edges indicate lower-than-average effective migration. Points indicate sampling localities and are colored by species assignment in model 3 (Fig. 5C). Within these three clades there is additional phylogeographic structure. Allohistium maydeni specimens from the Big South Fork resolve in a clade that is reciprocally monophyletic relative to all other Cumberland River basin populations (Fig. 3B). Similarly, specimens of A. cinereum from the Elk River resolve as the sister lineage of populations from tributaries in the upper Tennessee River. However, there is poor to moderate support for the placement of upper Tennessee River populations in the Clinch and Emory rivers. Bootstrap support for this node ranges from 66% to 84% (Supplementary Fig. S1). There is little phylogeographic structure within A. anas in the Duck River system, although the only specimen from the Buffalo River resolves as sister to all specimens sampled from the Duck River (Fig. 3B). Additional genetic samples from the Buffalo River would be required to confirm this slight population differentiation. Although to our knowledge we sampled all extant populations of Allohistium, the discovery of new populations could alter our phylogeographic conclusions. Population Genetic Patterns We examined patterns of population structure and effective migration within Allohistium. Imputed and unimputed data sets produced nearly identical results (Supplementary Figs. S2 and S3); therefore, we discuss only the results of the analyses with imputed genotypes. First, we estimated individual ancestry coefficients across a range of ancestral population (K) values. Cross-entropy scores exhibit a slight elbow at K = 5 (Supplementary Fig. S2). At K = 3, the three genetic clusters perfectly match the three main clades resolved in the maximum likelihood tree (Fig. 3B). All individuals are strongly assigned to a single genetic cluster with minimal admixture (minimum ancestry coefficient = 87.2%, average = 97.8%). At K = 4, specimens of A. cinereum from the Little River and the Elk River are placed into separate genetic clusters, and individuals from the Emory and Clinch rivers exhibit an admixed ancestry. At K = 5, specimens of A. maydeni are distributed among two clusters comprised of populations from the Big South Fork versus all other populations distributed in tributaries of the Cumberland River (Buck Creek, Obey River, and Rockcastle River), with minimal admixture between the two clusters (Fig. 3B). The effective migration surface estimated by FEEMS reveals lower-than-average migration edges isolating each of the three major Allohistium lineages (Fig. 3C). In the migration network, this manifests as a region of low migration edges near the confluence of the Duck River and Tennessee River. Populations of A. maydeni within the Cumberland River system and A. anas in the Duck River system are generally connected by higher-than-average effective migration edges. However, FEEMS identifies lower-than-average effective migration near the confluence of the Big South Fork and the Cumberland River proper, reflective of the K = 5 population structure within A. maydeni (Fig. 3B). Within the range of A. cinereum, we observe several regions of lower-than-average effective migration (Fig. 3C). This result is in agreement with the more complex phylogeographic and population structure identified for A. cinereum (Fig. 3A, B). Demographic analyses reveal an even more detailed perspective on the evolutionary history of the five Allohistium populations identified by the K = 5 population structure analyses. The crown age of Allohistium is estimated to be approximately 2.8 Ma, with subsequent divergence between A. maydeni and A. anas occurring approximately 1.8 Ma ( Fig. 4A). In contrast, intraspecific divergence is much more recent; the A. cinereum populations diverged approximately 312,000 years ago and the A. maydeni populations diverged only approximately 7000 years ago (Fig. 4A). Although ancestral effective population sizes (Ne) numbered in the millions, extant populations are much smaller, with some estimates as low as Ne = approximately 6000 individuals (Fig. 4A). Estimated migration rates are generally low between all population pairs, with no migration rates greater than 0.8 individuals per generation (Nm) (Fig. 4B). The highest migration rate estimates are from A. anas to A. maydeni in the Cumberland River basin (Nm = 0.799) and from A. cinereum in the Elk River to A. cinereum in the upper Tennessee River basin (Nm = 0.785). Best-fit demographic model for the Allohistium species complex. (A) History of lineage splitting and effective populations sizes, assuming a generation time of 2 years. Note that time is shown on a log scale. Allohistium populations are colored to match the K = 5 population structure in Fig. 3B. (B) The estimated asymmetric migration rate matrix. Migration rates are shown as the number of migrants per generation (Nm), whereas heatmap colors are scaled by log(Nm) for visual clarity. Phylogenomic Species Delimitation We examined support for alternative species delimitation scenarios in Allohistium using the GDI. This metric, bounded by 0 and 1, indicates the degree of genetic distinctiveness of a lineage. Prior comparative studies have developed approximate GDI thresholds for populations (<0.2) versus species (>0.7). Lineages with GDI values between 0.2 and 0.7 are considered to have an ambiguous status, where delimitation of species may rely on other lines of information such as phenotypic and ecological disparity (Jackson et al. 2017; Leaché et al. 2019). We estimated GDI for eight lineages (Fig. 5A). These lineages are the product of all possible splitting and lumping scenarios of five lineages in the maximum likelihood phylogeny: the population of A. maydeni from the Big South Fork, A. maydeni from other Cumberland River tributaries, A. anas from the Duck River system, A. cinereum from upper Tennessee River tributaries, and A. cinereum from the Elk River. Allohistium anas and A. maydeni both have mean GDI values greater than the 0.7 species-level threshold (Fig. 5B). All other lineages have mean GDI values in the ambiguous delimitation zone. Genetic and morphological delimitation support a three-species model for the Allohistium complex. (A) Branch numbers indicate the crown groups for eight possible species consistent with the inferred topology. (B) Mean and 95% credible intervals for the estimated genealogical divergence indices (GDI) of the eight possible species. (C) Species delimitation models for the morphological delimitation analyses. (D) CLUMPP H′ scores for each delimitation model appearing in panel C with different percentages of retained cumulative variance for the discriminant analyses. (E–H) Results from analyses with 80% retained cumulative variance. (E) Difference between observed model H′ and the mean permuted H′. Asterisks indicate models with significantly higher H′ than the permuted models. (F) Bar plot of individual assignments for M3. Colors represent assignment probability to a k-means cluster. (G) First two discriminant functions for M3. Points and convex hulls colored by species. Point shapes by k-means cluster assignment. (H) Meristic trait contributions to the first two discriminant axes for M3. Red indicates a negative contribution, and black indicates a positive contribution. LL = lateral line scales, PoreLL = pored lateral line scales, AbLL = scale rows above the lateral line, BlwLL = scale rows below the lateral line, Trans = transverse scale rows, CD = scales around the caudal peduncle, D1 = dorsal fin spines, D2 = dorsal fin rays, A1 = anal fin spines, A2 = anal fin rays, P1 = pectoral fin rays, Nape = % squamation of the nape, Cheek = % squamation of the cheek, Opercle = % squamation of the opercle, Breast = % squamation of the breast, Belly = % squamation of the belly. Morphological Species Delimitation We developed an unsupervised machine learning approach to perform clustering of individuals, assess trait diagnosability, and test alternative species delimitation models using multivariate phenotypic data sets. We deployed this method on a meristic trait data set for Allohistium comprised of 16 traits for 259 individuals (Supplementary Table S3). First, we assessed the fit of six delimitation models ranging from two to five species (Fig. 5C). For each model, we calculated H’, a pairwise similarity score ranging between 0 and 1, where larger values indicate a closer match between the groups identified by unguided clustering and the groups in the delimitation model. The H’ scores are not sensitive to the cumulative percentage of variance retained for the discriminant analyses (Fig. 5D). All delimitation models except M2 have statistically significant H’ scores based on permutation tests (Fig. 5E). M3 has the highest absolute H’ score and the largest difference between the observed H’ score and the mean permuted H’ score. The M3 species delimitation is consistent with the three major clades resolved in the maximum likelihood phylogeny that correspond to Allohistium anas, A. cinereum, and A. maydeni (Fig. 3B). We focus our examination of unguided phenotypic clustering results for M3, the best-fit delimitation model based on H’ scores. Most individuals are separated into the three phenotypic clusters; 89% of individuals have greater than 70% assignment probability to a phenotypic cluster (Fig. 5F). The three phenotypic clusters generally match the three species delimited in M3. Among individual specimens with greater than 70% assignment probability, two of the phenotypic clusters have high lineage specificity (Fig. 5F; Supplementary Table S7): 95% (86/91) of the individuals assigned to one phenotypic cluster are specimens of Allohistium maydeni and 86% (49/57) of the individuals assigned to another phenotypic cluster are specimens of A. cinereum (Fig. 5F; Supplementary Table S7). The third phenotypic cluster does not exhibit high lineage specificity: 60% (49/81) of the specimens in this phenotypic cluster are A. anas from the Duck River and 33% (27/81) are specimens of A. cinereum from the Elk River (Fig. 5F; Supplementary Table S7). The lack of strong lineage specificity for this phenotypic cluster is driven by geographic variation in phenotypic traits within A. cinereum. For instance, 67% (12/18) of A. cinereum individuals from the Elk River are assigned to this mixed phenotypic cluster (Fig. 5F). Our delimitation analysis identifies several meristic traits that load heavily on the discriminant functions separating the phenotypic clusters. The first discriminant function largely separates Allohistium anas from A. maydeni, with the cluster comprising specimens of A. cinereum occupying an intermediate position (Fig. 5G; Supplementary Fig. S4). Traits with the largest contribution to the first discriminant function include nape, cheek, and opercle squamation, as well as the numbers of first dorsal fin rays and number of scales around the caudal peduncle (Fig. 5H). The second discriminant function largely separates A. cinereum from A. maydeni, with the cluster comprising specimens of A. anas occupying an intermediate position (Fig. 5G; Supplementary Fig. S4). Traits with the largest contribution to the second discriminant function include the numbers second dorsal fin spines, first dorsal fin rays, transverse scale rows, and scale rows below the lateral line (Fig. 5H). To assess the performance of our phenotypic clustering and delimitation method, we analyzed a set of meristic trait data sets for 13 other darter species pairs (Ghezelayagh et al. 2025). These species pairs encompass a broad range of genetic divergence; sister species divergence times range from 83,000 to 8.14 Ma (Near et al. 2011). H’ scores for these species pairs range between 0.35 and 0.87 (Fig. 6). Three species pairs did not have significant H’ scores based on permutation tests. Notably, one of the statistically non-significant pairs is Percina tanasi and P. uranidea, which are taxonomically over-split and are not reciprocally monophyletic (Ghezelayagh et al. 2025). There is a significant positive correlation (R2adj = 0.3, P = 0.031) between H’ and the pairwise morphological Mahalanobis distance (Supplementary Fig. S5). Species pairs with higher H’ scores separate more strongly along the first PC axis and have phenotypic clusters with higher species specificity (Fig. 6). With an H’ score of 0.57, our three-species delimitation model for Allohistium is firmly nested within the range of H’ scores for other darter species pairs. Phenotypic delimitation results for meristic data from 13 darter species pairs presented in Ghezelayagh et al. (2025). Top left shows H’ values for each pair, with abbreviated species names. Asterisks indicate statistical significance based on permutation tests. Allohistium delimitation models are highlighted in white. Scatter plots show the first two principal component axes for each species pair, colored by species. Bar plots show individual posterior assignment probabilities and vertical black lines indicate breaks between the species. Discussion Although comparative morphological data are often not incorporated into molecular species delimitation analyses, phenotypic characteristics remain fundamental to biodiversity discovery and formal species descriptions. We introduce a novel approach to phenotypic species delimitation that operates without prior guidance. This method is applied to the Allohistium species complex, a group of freshwater fish inhabiting the Tennessee and Cumberland River systems in southeastern North America. To evaluate our method's effectiveness, we employ a comparative data set comprising 13 sister species pairs of darters. Our phenotypic analysis of Allohistium yields an optimal three-species model, aligning with results from our phylogenomic species delimitation analyses. We argue, as have others (Dayrat 2005; Padial et al. 2010), that species delimitation should always examine multiple lines of evidence. Our method enables explicit testing of species boundaries using morphological data. Three Genetically and Phenotypically Distinct Species of Allohistium Our phylogenomic analyses delimit three species of Allohistium: Allohistium cinereum (Storer, 1845), Ashy Darter, distributed in the Elk River and upper Tennessee River system, Allohistium maydeni Powers and Kuhajda, Redlips Darter, distributed in the Cumberland River basin, and Allohistium anas described in this study, Cinder Darter, distributed in the Duck River system. Phylogenomic and mtDNA analyses consistently resolve each of the three species as reciprocally monophyletic lineages (Figs. 2 and 3 ; Powers et al. 2004, 2012). Landscape genetic analyses reveal that gene flow is limited by valleys of low effective migration separating the three species (Fig. 3C). Likewise, demographic analyses show low migration rates between all species (Fig. 4B). Allohistium anas and A. maydeni have mean GDI values well above the “species” cutoff of 0.7 (Fig. 5B), but A. cinereum has a lower average GDI of 0.43. Given the phylogenetic relationships of Allohistium (Fig. 3B) and the rules of zoological nomenclature, if A. maydeni and A. anas are recognized as species, then A. cinereum must also be recognized as a species. We note that although the GDI “species” threshold of 0.7 is based on extensive meta-analysis of molecular data sets from many taxonomic groups (Jackson et al. 2017), it is still a heuristic and should be used with caution (e.g., Huang 2021). Thus, it is desirable to compare the results of the GDI heuristic approach with other, independent lines of evidence. Demographic analyses also offer support for the three-species model. The stem lineage of Allohistium cinereum originated approximately 2.8 Ma (Fig. 4A). Allohistium anas then diverged from A. maydeni approximately 1.8 Ma (Fig. 4A). Such divergence times are consistent with previous phylogenetic analyses, which showed that most described darter sister species are more than 1 Ma divergent (Near et al. 2011; Kim et al. 2023; MacGuigan, Orr, et al. 2023; Stokes et al. 2023). In contrast, A. cinereum populations diverged within the past approximately 300,000 years and A. maydeni populations only split in the last approximately 7000 years (Fig. 4A). There are a few darter sister species with similarly recent divergence times. For instance, E. perlongum and E. maculaticeps split within the past 11,000–33,000 years, but as a result of strong ecological selection between lake and river environments (Hubbs and Raney 1946; MacGuigan, Krabbenhoft, et al. 2023). Percina freemanorum and P. kusha share common ancestry approximately 350,000 years ago (Near et al. 2011), but their recent divergence is accompanied by substantial disparity in meristic trait distributions (Near et al. 2021). In these cases, additional lines of evidence were necessary to justify the delimitation and description of the species. Using our new phenotypic species delimitation method, we tested the fit of six different models, ranging from two to five species of Allohistium. All models were based on our inferred phylogenetic tree topology (Fig. 3B) except for M1, which is based on the current taxonomy of two species, where A. cinereum sensu lato is paraphyletic relative to A. maydeni (Fig. 5C). The three-species delimitation model (M3) is the closest fit to the unsupervised phenotypic clustering results (Fig. 5D). The three-species model also has the largest difference between the fit of the observed data set and the permuted data sets (Fig. 5E). This three-species phenotypic delimitation model is the same model best supported by phylogenomic analyses. Under the three-species model, discriminant analysis of the phenotypic data does not result in perfect assignment of individuals to their respective species (Fig. 5F, G). Two of the three phenotypic clusters exhibit high specificity for a single lineage (Fig. 5F; Supplementary Table S7). Yet individuals of Allohistium cinereum sensu stricto were split in their assignment between two different clusters. This indicates that many of the underlying meristic trait distributions have substantial overlap (Supplementary Figs. S6 and S7), but our phenotype delimitation method provides an objective strategy to identify traits that can diagnose species in a formal taxonomic description. For example, nape, cheek, and opercle squamation contribute heavily to the first discriminant function, which separates A. anas and A. maydeni (Fig. 5H). These trait distributions differ substantially between the two species (Supplementary Fig. S7). These differences are especially noticeable for cheek squamation, where 58% of A. anas specimens examined have more than 80% cheek squamation, whereas 64% of A. maydeni specimens have less than 20% cheek squamation (Supplementary Fig. S7). Likewise, first dorsal fin spines contribute heavily to the second discriminant function, which separates A. cinereum from A. maydeni (Fig. 5H), 81% of A. maydeni specimens examined have 11 or fewer first dorsal spines, whereas 76% of A. cinereum specimens exampled have 12 or more first dorsal spines (Supplementary Fig. S6). Therefore, although the three phenotypic clusters do not perfectly correspond to the three Allohistium species, our approach offers a foundation for further investigation and species description. Our results indicate strong multivariate diagnosability of A. maydeni. Although A. cinereum and A. anas exhibit greater morphological overlap relative to A. maydeni (Supplementary Figs. S6 and S7), certain meristic traits still exhibit differences that help distinguish these species. Conservation Genetics of the Allohistium Species Complex Our phylogeographic, population structure, and demographic analyses reveal differences in population connectivity and genetic diversity within the Allohistium species complex. All estimated migration rates between populations are less than one migrant per generation, a gene flow threshold below which populations are likely to diverge via selection or genetic drift (Mills and Allendorf 1996; Wang 2004; Nathan et al. 2017). However, Allohistium anas and A. maydeni exhibit shallower phylogeographic structure and reduced genetic diversity compared to the A. cinereum populations (Fig. 3). Demographic analyses also show that population structure within A. cinereum is relatively old, dating back approximately 300,000 years (Fig. 4A). Landscape genetic analyses reveal high population connectivity within A. anas and A. maydeni compared to the low effective migration edges isolating A. cinereum in the Elk River and upper Tennessee River system (Fig. 3C). Together, this suggests that there are strong dispersal barriers between the Elk River and upper Tennessee River. This pattern is strikingly similar to other freshwater fishes distributed in the Tennessee and Cumberland basins. For example, in the Etheostoma kennicotti species complex, the Cumberland River basin comprises at most three genetic clusters, whereas the upper Tennessee River basin can be subdivided into as many as 10 genetically distinct clusters (Near et al. 2023). These population-level patterns are also reflected in the higher number of fish species endemic to the Tennessee River system versus the Cumberland River system (Warren and Burr 1994). Thus, the population-level processes reducing effective migration and increasing genetic diversity among geographically disjunct populations of A. cinereum are reflective of the broader evolutionary processes generating freshwater biodiversity. Historically, Allohistium likely had a more contiguous range within the Tennessee and Cumberland rivers (Shepard and Burr 1984; Powers et al. 2004, 2012). The large ancestral effective population sizes in our demographic analysis (Fig. 4A) may reflect this broader historical distribution. However, impoundment of these rivers heavily altered available habitats and may have caused local extirpations of Allohistium (Etnier and Starnes 1993). There have been substantial conservation efforts to improve the resiliency of this species complex, including range-wide surveys, captive propagation, and reintroduction (https://ecos.fws.gov/ServCat/DownloadFile/171240). Although A. maydeni in the Cumberland River system and A. anas in the Duck River system are genetically homogenous, future conservation efforts for A. cinereum in the middle and upper Tennessee River basin should carefully consider population genetic structure. For instance, the Little River was selected as the source population for reintroduction into the Tellico River from 2012 to 2016 due to the close proximity of these river basins (https://ecos.fws.gov/ServCat/DownloadFile/171240). However, our results show that the Little River population of A. cinereum is genetically distinct from other nearby populations in the upper Tennessee River basin. This is exhibited by a long internal branch leading to the Little River specimens and a distinct Little River genetic cluster at K > 4 in our population structure analyses (Fig. 3B). Given that the historic Tellico River population was likely extirpated, stocking individuals from the Little River would have no long-term consequences. Nevertheless, future efforts should account for the genetic connectivity of populations in the Tennessee River to limit potential disruption of local adaptation via admixture. The population structure and patterns of admixture among A. cinereum populations merit additional study with increased sampling resolution. We also identified very low effective population sizes (approximately 12,000 total individuals) for A. maydeni, particularly striking when compared to the much larger effective population sizes of A. cinereum (Ne = approximately 240,000) and A. anas (Ne = approximately 170,000) (Fig. 4A). However, we caution that demographic parameters, including effective population size, are point estimates reliant on numerous assumptions such as generation time and mutation rate. Therefore, the absolute values of these parameters should be interpreted with caution. Nevertheless, the relative disparity in effective population sizes suggests that A. maydeni warrants further study to assess its genetic diversity and conservation status. Comparative Species Delimitation and the Importance of Phenotype To place our delimitation of Allohistium in a reference-based comparative framework and test the efficacy of our new method, we analyzed phenotypic data sets for 13 darter sister species pairs. The model-fit metric H’ for the three-species Allohistium model falls neatly within the range of H’ values for recognized darter species, ranking 6th highest (Fig. 6). Therefore, the delimitation of three species of Allohistium exhibits phenotypic disparity and diagnosability that is consistent with other long-recognized species of darters. The Allohistium delimitation models with the next highest H’ values, models 1 and 5 (Fig. 5C), rank only 10th and 11th highest, closer to the bottom end of the distribution of H’ values in the reference taxonomic data set. Thus, this explicit comparative framework provides another line of evidence supporting our three-species delimitation of Allohistium. Three of the darter sister species in our comparative data set have H’ values that are not statistically significant based on permutation tests. One of the species pairs with a statistically insignificant H’ values is Percina tanasi Etnier and P. uranidea (Jordan and Gilbert). These species were recently the subject of a rigorous species delimitation study, which found little genetic or phenotypic differentiation that resulted in the synonymization of P. tanasi with P. uranidea (Ghezelayagh et al. 2025). Our unguided phenotypic delimitation analyses agree with the assessment that P. tanasi should not be recognized as a distinct species. Species pairs of darters with low, statistically insignificant H’ values demonstrate common challenges in phenotypic species delimitation. For instance, the PCA of meristic data for Nothonotus bellus and N. camurus shows considerable overlap in their morphospace, resulting in poor diagnosability (Fig. 6). However, this overlap does not mean these species lack phenotypic distinction. Our data set lacks key traits used to describe N. bellus, such as color patterns, nape squamation degree, and snout shape (Zorach 1968), whereas it was explicitly noted that meristic differentiation between these species is less pronounced. This underscores a limitation of all phenotype-based species delimitation approaches: there may always be additional, unquantified phenotypic traits crucial for species delimitation and diagnosability. Fortunately, our method is adaptable to this challenge. It can incorporate various data types in a single analysis and easily scales to accommodate even the most extensive phenotypic data sets. Like Nothonotus bellus and N. camurus, the sister species pair N. denoncourti and N. tippecanoe have a low H’ value indicating low phenotypic divergence (Fig. 6). Unlike N. bellus and N. camurus, the taxonomic description of N. denoncourti identifies several meristic traits with modal differences between the species, including the numbers of lateral line scales, pored lateral line scales, and dorsal fin spines (Stauffer and van Snik 1997). All of these traits are included in our data set. We can observe that the two species exhibit a degree of separation in the meristic morphospace, but this separation occurs along the second PC axis (Fig. 6). Therefore, it appears some unknown variable explains more of the phenotypic variation (i.e., PC1) than species status (i.e., PC2). Qualitatively, this unknown variable does not appear to be consistent with either geography or sex (Supplementary Fig. S8). This result does not mean that N. denoncourti and N. tippecanoe are not phenotypically diagnosable, but it does suggest that intraspecific variation in the sampled meristic traits within these species is greater than interspecific variation. Our phenotypic delimitation method, as exemplified by the case of Nothonotus denoncourti and N. tippecanoe, prioritizes diagnosability over overall morphological differentiation. Although greater inter-species differentiation often indicates higher diagnosability, this is not always the case. We compared H’ values (a diagnosability metric) with Mahalanobis distances (an absolute differentiation metric) for each species pair (Ghezelayagh et al. 2025). These metrics showed a moderate, positive linear relationship (Supplementary Fig. S5). Some species pairs with high Mahalanobis distances exhibited only moderately high H’ values. For instance, Percina macrocephala and P. williamsi had the highest Mahalanobis distance (2.38) but only the 7th highest H’ (0.53). These metrics address different aspects of phenotypic divergence: Mahalanobis distance assesses the degree of morphological disparity between a priori species; the traditional approach to phenotypic species description. In contrast, our method evaluates how well phenotypic clusters in the data align with a priori delimitation models. We argue that robust species delimitation should emphasize phenotypic diagnosability over mere phenotypic distinctiveness. A broader discussion of phenotypic species delimitation methods should consider their ultimate goals and limitations. For extant organisms, species delimitation should ideally incorporate both genetic and phenotypic evidence to identify distinct, independent evolutionary lineages, as relying on one line of evidence can lead to erroneous delimitations. Although it has allowed for the discovery of numerous cryptic species, genetic delimitation alone may not sufficiently distinguish between population structure and species-level diversity. The aim is typically to identify independent evolutionary lineages that are both genetically and morphologically distinguishable. Many have argued for combining genetic and phenotypic data into a single delimitation analysis (Dayrat 2005; Edwards and Knowles 2014; Karbstein et al. 2024). In that sense, our method is not truly integrative. The original DAPC approach was developed for genetic markers, so our method could theoretically be extended to incorporate both morphological and genetic data. However, there are two reasons this may be undesirable. First, DAPC has no underlying evolutionary model; delimitation with genetic data is likely better served by a biologically grounded model such as the multispecies coalescent. Second, genetic and morphological data often disagree about species boundaries. There may be strong genetic differentiation with little or no morphological divergence (cryptic species). The opposite can also be true, with strong morphological differences despite low levels of genetic divergence (e.g., Poelstra et al. 2014). For this reason, we believe it is advantageous to examine molecular and morphological evidence separately when delimiting species. If there is incongruence, the investigator must make a judgment call or consider other lines of evidence such as coloration, behavior, or reproductive isolation. Morphology serves important roles in species delimitation, both theoretically as a proxy for neutral evolutionary drift, ecological selection, and sexual selection—processes driving reproductive isolation—and practically for species identification in field work, conservation efforts, and ecological studies. It is important to note that traits or trait combinations are rarely 100% diagnostic, with species often having overlapping trait distributions. Nevertheless, these traits can still be valuable in defining species. In contrast to classic taxonomic approaches that look for mean or modal differences between predefined groups, our PhenoDelimit method facilitates objective identification of informative traits without a priori knowledge. Supplementary Material