Sampling, DNA extraction and sequencing

Sixty-four muscle tissue samples of P. grosskopfii and fifty-seven of P. yuma were collected between 2014 and 2021 and provided by Integral S.A., through two scientific cooperation agreements (19th September 2013; Grant CT-2013–002443), by Instituto Humboldt, and less than 10% obtained under collection permit ANLA 1467 de 2019. All samples were collected at 24 sites along ~ 460 km and ~ 1200 km in the mainstream and floodplain lakes of the upper, middle and lower sections of the Colombian Cauca River and Magdalena River, respectively (Fig 1; S1 Table), during the dry season, a period that facilitates capture by traditional fishing gear. DNA was extracted from ethanol-preserved tissues by using the GeneJET Genomic DNA Purification Kit (Thermo Scientific). DNA quality was evaluated by electrophoresis on 0.8% agarose gels and EZ-vision® staining (AMRESCO). DNA was quantified with a NanoDrop 2000 spectrophotometer (Thermo Scientific) and diluted to a final concentration ranging from 30 to 50 ng/uL.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Sampling sites along ~1600 linear kilometers of the upper, middle, and lower sections of the Cauca and Magdalena rivers, including floodplain lakes and wetlands.

Gray circles represent Pimelodus grosskopfii and black circles represent Pimelodus yuma. Bicolored circles indicate sites where individuals of both species were sampled. River sections are color-coded as follows: red for Upper Cauca, light blue for Middle Cauca, green for Lower Cauca, pink for Upper Magdalena, indigo blue for Middle Magdalena, and orange for Lower Magdalena. River network data were obtained from Rivers in Colombia (World Bank Data Catalog; public domain), licensed under Creative Commons Attribution 4.0 (CC BY 4.0), available at https://wbwaterdata.org/dataset/rivers-colombia. Background raster data were obtained from Natural Earth (public domain), available at http://www.naturalearthdata.com/. The map was constructed using these data sources in the R environment (https://www.r-project.org/).

https://doi.org/10.1371/journal.pone.0351301.g001

Both RAD library preparation and sequencing were carried out by the National Center of Genomic Sequencing (Colombia) in collaboration with Floragenex Inc. (USA), following their internal protocols (see https://floragenex.com for more details). At least 0.5 μg of high-molecular-weight DNA per individual was used for RAD library preparation. Genomic DNA was digested using the 6-bp restriction enzyme PstI, followed by fragmentation via sonication. Libraries were constructed through adapter ligation, PCR amplification, and size selection targeting fragments within a range of approximately 200–300 bp, following proprietary internal protocols for RAD-seq library preparation of Floragenex. Sequencing was performed on a single lane of an Illumina HiSeq 2000 machine, generating single end reads of ~125 bp. This sequencing configuration was expected to produce approximately 80 million reads per lane, yielding at least 0.5–1 million reads per individual and an average sequencing depth per locus of approximately 10–30 × , depending on library complexity and quality.

It is important to clarify that all biological material used in this study consisted exclusively of tissue samples provided by third parties through institutional agreements described above. In addition, DNA samples and tissue remnants are available in the Molecular Biology Laboratory at the Universidad Nacional de Colombia, Medellín campus (see S1 Table).

Genomic data processing

After sequencing, the genomic data quality was analyzed in FASTQC (Bioinformatics Group at the Babraham Institute, UK) [24]. The samples were demultiplexed using cutadapt [25], and SNPs were extracted using DiscoSnpRad [26] with default parameters except that presence indels and up to five SNPs per locus were permitted for construction of the loci. The resulting Variant Call Format (VCF) was quality filtered considering a Phred score  >  20, it was removed variants detected in 20% and less populations, set minimal read depth of seven per allele and finally, filtered to eliminate potential paralogues (SNP rank  <  0.4) [23]. This SNP rank, implemented in DiscoSnpRad [26], is a score ranging from 0 to 1 and reflects the confidence that a variant corresponds to a true bi-allelic polymorphism rather than arising from sequencing errors or collapsed paralogous regions. The SNP rank is computed from local graph properties in the de Bruijn graph, including coverage balance between alleles and the structural consistency of variant-associated bubbles. Low SNP rank values are typically associated with variants showing abnormal coverage patterns or complex graph structures, which are indicative of potential paralogous sequences or repetitive regions. Therefore, we applied a conservative threshold and removed all variants with SNP rank < 0.4, retaining only high-confidence SNPs for downstream analyses.

To generate the final genetic dataset used for population genetic downstream analyses (e.g. SNPs genotype matrix in VCF, STRUCTURE, GENPOP or ARLEQUIN format), we used the component populations.pl implemented in the software STACKS v.1.2 [27] using as inputfile the resulting filtered VCF above generated by DiscoSnpRad. The filtering parameters in STACKS included that each individual was considered a population in the map file, then the minimum number of populations (-p) that a locus must be present in to process it was set to 90% of the total sample (10% of missing data). Other filtering values for “-p”, tolerating ∼0% and ∼ 5% of missing data in the final dataset were also explored according to Martínez et al (2022) [28], however, these last two datasets showed less statistical power to discriminate samples (in terms of explained variance) and were then discarded for downstream analysis. Likewise, in STACKS, the minimum percentage of individuals in a population required to process a locus for that population was set to 100% (−r = 1), as well as a minor allele frequency (MAF) of 0.01 and retained only one SNP per locus. All other parameters for the analysis were kept as default settings.

Considering that statistical analysis that seek to infer gene flow, structure and diversity indices are based on the principles of neutral variation, we decided to examine the existence of outlier loci to remove them from the final dataset. For this, an analysis was performed with BAYESCAN v.2.1 software [29], which identifies outliers loci using differences in allele frequencies between populations with the best efficiency when compared with alternative programs to detect outlier loci [30]. For our analysis, parameters for the BAYESCAN analyses included 10:1 prior odds for the neutral model, and 20 pilot runs consisting of 5,000 iterations each, followed by 250,000 iterations with a burn-in length of 50,000 iterations. The loci were ranked according to their estimated posterior probability, being classified as putative outlier when ≥0.95 [31].

Finally, after removing outlier loci, the neutral dataset was evaluated for linkage disequilibrium using the standardized index of association (rbarD) implemented in the poppr package [32]. This statistic measures the overall non-random association of alleles across loci (i.e., multilocus linkage disequilibrium), allowing assessment of the degree of non-independence among markers. Unlike pairwise measures such as the squared allelic correlation (R²), which quantify linkage between specific pairs of loci, rbarD provides a global estimate of linkage disequilibrium (LD) across the dataset [32]. In this study, a threshold of rbarD > 0.2 was used as a conservative criterion to identify potential non-independence among loci and reduce redundancy in the SNP matrix, following Escobar et al. (2024) [23].

The RADseq raw demultiplexed dataset has been submitted to NCBI as Bioproject PRJNA1217604.

Population structure analysis

For the purposes of geographic population analysis, samples of both P. grosskopfii and P. yuma were grouped according to natural longitudinal sections commonly recognized for the Cauca River (Upper Cauca, Middle Cauca, and Lower Cauca) and the Magdalena River (Upper Magdalena, Middle Magdalena, and Lower Magdalena) [33,34]. However, since the Lower Cauca and Lower Magdalena regions are geographically and bioecologically adjacent and given the reduced individual sample sizes in these areas for this study, they were combined to ensure statistical robustness, being collectively designated as samples from the Lower Cauca-Magdalena Basin, hereafter referred to as ‘Lower Cauca’ for graphical purposes. In the end, we defined five and three geographic groups for all subsequent analyses, corresponding to P. grosskopfii and P. yuma, respectively.

To quantify and evaluate the genetic structure and the degree of connectivity of the populations, multiple statistical approaches were used. The first approach was the Bayesian one, which was implemented in STRUCTURE v.2.3.4 software [35] under the admixture model and correlated allel frequencies models. For each possible number of groups (K  =  1–8 for P. grosskopfii; K = 1–6 for P. yuma), 10 independent runs for 1,000,000 Markov chain Monte Carlo iterations each were performed, discarding 200,000 runs as burn-in. The optimal number of genetic clusters (K) was inferred using the Puechmaille (2016) estimators, which builds upon earlier metrics and provides enhanced visualization tools [36–38]. We applied this method through the StructureSelector web server [39] because it outperforms traditional approaches by more reliably detecting the correct population structure under conditions of uneven or unbalanced sampling. New estimators included the median of medians (MedMed), the maximum medians (MaxMed), the median of means (MedMean), and the maximum of means number of population clusters (MaxMean). Next, we implemented the multivariate statistics of principal component analysis (PCA), using the R package ADE4 [40], ADEGENET [41], and FACTOEXTRA [42]. The PCA is a dimensionality reduction method used to transform and project data points onto fewer orthogonal axes that can explain the greatest amount of variance in the data, allowing for help in population delimitation, without selecting geographic groups a priori to force their grouping or maximize their variances. Likewise, genetic differentiation among geographic sections, and genetic groups suggested by STRUCTURE, was calculated using the standardized statistics Fixation Index (F_ST) [43] using ARLEQUIN v.3.5.2 [44]. Both PCA and F_ST were performed with neutral, outlier and all SNPs, to identify whether there was any differentiation in the grouping or contribution and strength of presumable outliers/adaptive loci on the structure patterns.

We defined putative genetic stocks for downstream analyses by integrating three lines of evidence, including co-ancestry inference with STRUCTURE, as well as genealogical relationships and historical demography, as described below.

Phylogeographic relationships

Additionally, a phylogeographical reconstruction analysis was conducted using the neutral, outliers, and combined datasets to provide genealogical support (evidence of reciprocal monophyly) for all the allele frequency-based results already inferred and to evaluate the historical (neutral) and recent (non-neutral) contributions to the observed patterns of genetic structure from a phylogenetic perspective [22]. For this purpose, a maximum likelihood (ML) tree was inferred in RAxML v.8.2.X [45] using the GTR + G evolutionary model, using each type of dataset (neutral, non-neutral, and all loci). For the full ML tree searches, 1000 replicates of the rapid bootstrap algorithm of RAxML were used to account for uncertainty in the estimation of the tree topology [46] in the CIPRES Science Gateway [47]. Phylogenetic trees were independently inferred for each dataset (e.g., neutral and outlier SNPs) or combining them. In RaxML trees the branch lengths represent the number of substitutions per site. Because each dataset differs in the number, composition, and distribution of variable sites, branch length scales are not directly comparable among trees.

Detecting selection and adaptive variation

To evaluate the existence and contribution of adaptive variation to the observed population structure patterns, a selection detection analysis was performed using BAYESCAN v2.1 [29], employing the same running parameters described above, but inferring selection through the False Discovery Rate (FDR) method. The FDR is defined as the expected proportion of false positives among putative outlier markers. For this, it was set a cutoff of 0.001 (0.1%) to classify loci under selection with high confidence. BAYESCAN outputs were visualized as log10(q-values), such that an FDR of 0.001 threshold corresponds to a value of –3 on the plotted scale. Likewise, positive alpha values were then used to distinguish SNPs under directional selection, whereas negative alpha values were used to detect purifying selection [29]. Although FDRs between 1.4% and 5% are commonly admitted as the percentage of false positives within the loci considered under selection in some studies (e.g., Lotterhos and Whitlock 2014) [48], we wanted to be as rigorous as possible in determining that our loci were under selection, using an FDR of 0.1% equivalent to 1/140–1/500 of the probability of occurrence of false positives used previously in other studies.

We also inferred evidence for selective sweep by estimating the Linkage Disequilibrium (LD) [49]; following the methodology of Escobar et al. (2024) [23]. We applied the standardized index of association (rbarD) for the neutral and adaptive loci datasets, using ‘poppr’ software [32]. The analysis was carried out under two hierarchical levels of genetic variation: including both differentiated stocks (all pops) in the analysis and within each stock or subpopulation, to allow, respectively, for testing the LD under a scenario of positive selection (all biological groups) vs. a non-selective scenario (individual biological groups). The principle of the analysis is that a selective sweep signature is a localized pattern of LD levels, characterized by high LD on each side of a beneficial mutation (adaptive loci) and low LD between loci that are located on different sides of the beneficial allele (loci under neutral variation) [49].

Genetic diversity and demographic inferences of biological groups (stocks)

The diversity parameters were quantified in terms of observed heterozygosity (Ho) and expected heterozygosity (He), nucleotide diversity (Nd) and Percentage of Polymorphic loci (%P) calculated in ARLEQUIN v.3.5.2 [35] with 100,000 permutations, using neutral loci. In addition, the inbreeding coefficient (F_IS) was calculated in the same program, under similar conditions. The diversity parameters calculations were inferred for each genetic stock observed through STRUCTURE and phylogenetic analyses, avoiding the Wahlund effect on our metrics. Geographical structure affects allele frequencies over space and consequently the proportions of different genotypes in the local populations. It leads to the Wahlund effect, which can be defined as the excess of homozygotes or the deficit in heterozygotes observed in a sample of individuals obtained from a structured population, even when the local populations are randomly mating [50]. The contemporary effective population size (Ne) was calculated by each genetic stock using NeEstimator v.2 [51], using the linkage disequilibrium method, assuming 0.02 and 95% confidence intervals of the Jackknifing. This method is particularly robust for population structures with migration rates of <0.1 [52] and hence panmictic populations. Because of this, it was not used to calculate effective population size of the geographic populations, but only for biological (stocks) groups.

In addition, according to the observed genetic structure patterns and biological groups formation through STRUCTURE and PCA analyses, we defined genetic populations/stocks, and then, we quantified the gene flow in terms of the effective number of migrants per generation (Nm) between them following the spirit of Sanchez-Bernal et al. (2023) [53] using the program Migrate v3.6.11 [54]. For this purpose, we tested different migrations models between the paired predefined populations (1:full migration (xxxx); 2:unidirectional (x0xx); 3:unidirectional (xx0x); 4:full isolation (x00x)), and then, we selected the best one by comparing their Marginal Likelihood based on the Bezier approximation score, through the Bayes Factor method [LBF = 2 (ln(Prob(D | Model1) – ln(Prob (D | Model2)]. For Theta and M parameters between populations, we used uniform priors. The search strategy used static heating scheme with four chains at different temperatures (1.0, 1.5, 3.0, 1000000.0) and four replicates of one long chain with 50 million generations, discarding a burn-in of 5 million and sampling every 500 generations. To verify the convergence of parameter estimates, we examined the autocorrelation (<0.2), the effective sample size (ESS > 500) and the correct distribution of the probability density plots for estimates of the parameters. To convert the scaled parameter “M” to biological data in terms of number of effective migrants (Nm), we assumed a conservative generation time of 2.5 years used by Martínez et al. (2022) [28] for both species from the Magdalena-Cauca basin. Then, we assumed a substitution rate of 5.0 × 10−9 mutations per site per year, derived from a geometric mean between the conservative nuclear rate of 1.0 × 10−9 mutations per site per year [55] and the nuclear genomic substitution rate of 2.5 × 10−8 mutations per site per year used for other catfish species [56].

We used the program BEAST v.1.8.4 [57] to investigate patterns of changes in effective population size throughout the coalescent time of the Pimelodus species, by using Bayesian Skyline Plot approach. All analyses were run for 50 million generations with a burn-in of 10% initial samples, sampling every 10,000 topologies. We used the HKY+GAMMA model of molecular evolution and a lognormal relaxed molecular clock. Priors were adjusted based on preliminary analyses. Final analyses were repeated two times to assure convergence of estimated parameters, and the independent runs were combined. The absolute time scale of the Bayesian skyline plot was calculated using a substitutions rate explained and justified above for the Migrate software.

Putative identification of flanking region of the outlier SNPs

Then, the complete RAD-seq loci (contig) containing the outlier SNPs were extracted from the “.filtered.fa” archive generated by DISCOSNPRAD software, previously identified by BAYESCAN, and submitted to BLAST to determine their identity putatively. Given that both Pimelodus species are non-model species (without a whole sequenced, assembled, or annotated genome), the only expectation was that our loci would match genomes of phylogenetically close species (catfishes) or, failing that, teleosts. It is understood that homologous genes between different species have suffered mutations in their evolutionary trajectory during the divergence, hence more than similarity, the probability of the expected value (E-value < 0.001) was considered as a criterion for their putative nomination. This estimates the expected number of random alignments with a particular score or better that could be found by chance in a given database search. In other words, it represents the likelihood that a specific sequence alignment is attributable to chancer rather than a true biological relationship between the sequences.

Additionally, we performed an enrichment analysis in the PANTHER V19.0 program [58,59] on its online platform. The analyses were focused solely on shedding light on the biological process to which the set of genes could be associated, those that were previously identified using BLAST for the flanking regions of the outlier SNPs. For this purpose, we implemented the Protocol Update for Large-scale genome and gene function analysis with PANTHER Classification System, developed by Mi et al. (2019) [60], using a generic ID list of genes, Danio rerio and Oryzias latipes as representative teleost models with well-annotated genomes, due to the limited availability of reference organisms; and selecting a functional classification viewed in graphic charts. Likewise, to facilitate the biological interpretation of molecular functions and biological processes identified using the PANTHER Classification System—which integrates curated annotations from the Gene Ontology database—we employed the Gene Set Enrichment Analyst (GSEA), an artificial intelligence–assisted approach using ChatGPT as an auxiliary interpretative tool. This approach was used exclusively to generate structured summaries of gene functions based on enrichment [61] (https://chatgpt.com/g/g-DDiYl6Gbd-gene-set-enrichment-analyst). The AI tool was not used for data generation, statistical analysis, or hypothesis testing, but solely as an aid for organizing and summarizing functionally annotated information previously inferred from PANTHER.

Genomic data processing

A total of 35.79 million unique reads were obtained for P. grosskopfii and 25.15 million for P. yuma, with an average size of 122 bp. Each individual had on average 559,354 reads for P. grosskopfii and 449,081 for P. yuma. After extracting SNPs and filtering the VCF by coverage, 55,599 and 118,587 SNPs were retained respectively for P. grosskopfii and P. yuma. The mean depth per SNP in each individual ranged between 12.78 and 6,364.41 for P. grosskopfii, and between 12.54 and 6,043.61 for P. yuma. Finally, after filtering by STACKS (population module), we retained 2,475 and 7,611 variable loci (representing equal number of SNPs), respectively for each species, with 10% of missing data both. Datasets with other percentages of missing data [P. grosskopfii: 0% (0 SNPs) and 5% (471 SNPs); P. yuma: 0% (5 SNPs) and 5% (2,374 SNPs)] showed null or fewer abilities for inferring population structure among groups than the 10% missing data SNPs-matrix for both species in terms of percentage of explained variance for the two first principal components (S2 and S3 Figs), and therefore, this last one matrix was selected for the downstream analyses [28]. Six individuals of P. grosskopfii and two of P. yuma were removed due to having an excessive amount of missing data (>70%), retaining 58 and 55 individuals, respectively, for subsequent analyses. Next, according to the BAYESCAN analysis for outlier detection (S4 Table; S5 Table), we removed 44 and 15 non-neutral SNPs, retaining 2,431 and 7,596 neutral SNPs, respectively, for P. grosskopfii and P. yuma. After reviewing for LD, according to the analysis generated by using the ia () function, which calculated the IA overall 2,431 and 7,596 neutral loci in the datasets, the rbarD was 0.1984302 for P. grosskopfii and 0.1113031 for P. yuma (both rbarD  <  0.2). Therefore, no SNPs were removed from the datasets, which were then used for subsequent population structure, demographic and diversity analyses.

Population genetic structure

The structure based on Bayesian analysis revealed two genetic clusters (K = 2) or populations for both species (Figs 2 and 3), which was confirmed by the Puechmaille method (S6 and S7 Figs). These two genetic populations coexist in all sampled sections along the rivers as confirmed by the co-ancestry histogram (Figs 2A and 3A). The structure and genetic differentiation are not shown to be associated with basin geography as confirmed by the PCA analysis (Figs 2B and 3B), but rather with the existence of two differentiated genetic groups (stocks) that, although sharing the same space, showed low genetic interaction (Figs 2C and 3C). This was corroborated by the F_ST values among geographical areas, which were shown to be non-significant in all cases, ranging between 0.00051–0.02105 for P. grosskopfii, and 0.00018–0.00365 for P. yuma (P > 0.05) (Table 1). Contrary, the F_ST between stocks was highly significant for both species, being 0.05677 for P. grosskopfii, and 0.10067 for P. yuma (P < 0.0001) (Table 1).

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. F_ST values between geographic areas and genetic stocks in the Magdalena-Cauca basin for Pimelodus grosskopfii and Pimelodus yuma, using neutral (n) and adaptive (o) loci.

https://doi.org/10.1371/journal.pone.0351301.t001

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Population structure analysis in Pimelodus grosskopfii.

(A) Coancestry histogram inferred using STRUCTURE for individuals sampled in Upper Cauca, Middle Cauca, Lower Cauca, Middle Magdalena, and Upper Magdalena. (B) Principal Component Analysis (PCA) showing the association between geography and population structure. (C) PCA showing the association between genetic stocks, as previously identified by STRUCTURE, and the observed population structure. Sample sizes per stock are as follows: Stock 1 (n = 39) and Stock 2 (n = 19).

https://doi.org/10.1371/journal.pone.0351301.g002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Population structure analysis in Pimelodus yuma.

(A) Coancestry histogram inferred using STRUCTURE for individuals sampled in Upper Cauca, Middle Cauca, Lower Cauca, Middle Magdalena, and Upper Magdalena. (B) Principal Component Analysis (PCA) showing the association between geography and population structure. (C) PCA showing the association between genetic stocks, as previously identified by STRUCTURE, and the observed population structure. Sample sizes per stock are as follows: Stock 1 (n = 46) and Stock 2 (n = 9).

https://doi.org/10.1371/journal.pone.0351301.g003

Finally, given that at least two of the three lines of evidence were concordant in delineating genetic stocks in P. grosskopfii and P. yuma, we infer that both species comprise two distinct stocks. This inference is supported by the evidence of historical divergence in evolutionary trajectories (neutral genealogies), reciprocal monophyly (neutral+outlier genealogies), distinct demographic histories (see below), and differentiated co-ancestry patterns, the latter being particularly defined only in P. grosskopfii.

Genealogical relationships

The phylogeographical approach based on the genealogical analyses performed with each dataset revealed different relationship patterns between stocks (Fig 4). When using the “neutral” or “neutral+outliers” datasets, it was consistently demonstrated that the putative stocks are not statistically reciprocal monophyletic groups (Fig 4A and 4B). However, a reciprocal monophyly pattern was observed between stocks when only the outlier loci were used, showing two well-supported groups (100%), each one represented by a genetic stock (Fig 4C). For both species, in the phylogenetic reconstruction involving neutral SNPs, the putative stock 2 was shown as a subpopulation derived from the stock 1, which was, in turn, the ancestral one or basal population.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Phylogeographic reconstruction analysis for Pimelodus grosskopfii and P. yuma in the Magdalena–Cauca basin, based on Maximum Likelihood inference using RAxML.

Analyses were performed using (A) neutral loci, (B) neutral + outliers loci, and (C) outlier loci only. Blue indicates Stock 1, and orange indicates Stock 2.

https://doi.org/10.1371/journal.pone.0351301.g004

Selection and adaptive variation

According to BAYESCAN, the selection was detected as an evolutionary force acting on the populations, at the very conservative value for FDR of 0.001 (0.1%) cut-off (Fig 5A and 5D), providing strong evidence consistent with its existence and strength. In addition, the type of selection signature was consistent with directional, according to the positive alpha values (S4 and S5 Tables) (negative values are associated with purifying selection). Likewise, the PCA, using “outlier” loci, showed a similar pattern of two well-defined stocks, with a percentage of explanation of the total variance for the first two components of 77.1% for P. grosskopfii, and 70.3% for P. yuma. However, when “neutral+outlier” loci datasets were included for the same analysis, the percentage of explanation of the variance decreased drastically to 13.1% and 13.4%, respectively for each species (Fig 5B-5C, Fig 5E-5F). A similar pattern was observed in the F_ST estimated between stocks using outliers loci, which showed a very strong and highly significant (P < 0.0001) fixation index of 0.51914 for P. grosskopfii, and 0.76289 for P. yuma, decreasing drastically to non-significant (P > 0.05) values of 0.06958 and 0.10486, respectively for each species (Table 1) when “neutral+outlier” loci datasets were included for the analysis.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Natural selection and adaptive variation analysis in Pimelodus grosskopfii and P. yuma from the Magdalena–Cauca basin.

(A) and (D) show BAYESCAN analyses of 2,475 and 7,611 single nucleotide polymorphisms (SNPs) for P. grosskopfii and P. yuma, respectively. The vertical axis represents locus-specific F_ST values, while the horizontal axis shows the logarithm of Q-values. The vertical black line indicates the threshold used for outlier detection (false discovery rate (FDR) = 0.001), represented as log10(q-value) = –3. Each dot corresponds to an individual SNP. Outlier SNPs under positive selection are located above the upper-right threshold line. (B) and (E) illustrate Principal Component Analyses (PCA) based on outlier loci for P. grosskopfii and P. yuma, respectively. (C) and (F) show PCA results using both neutral and outlier loci for P. grosskopfii and P. yuma, respectively.

https://doi.org/10.1371/journal.pone.0351301.g005

As a selective sweep signature, the LD, inferred through the standardized index of association (rbarD), showed, as expected under a genomic selective sweep scenario, a high LD for both P. grosskopfii (rbarD = 0.6781) and for P. yuma (rbarD = 0.6050) on each side of a beneficial mutation (adaptive/outliers loci); and a low LD for both [P. grosskopfii (rbarD = 0.0212); P. yuma (rbarD = 0.0746)] between loci that are located on different sides of the beneficial allele (loci under neutral variation) (Table 2). However, when each population was evaluated separately [removing the effect of differential environments or adaptive factors (directional selection bias)], the LD was very low in both the stock 1 [rbarD (neutral) = 0.0158–0.0546; rbarD (adaptive) = 0.1256–0.1651] and stock 2 [rbarD (neutral) = 0.0032–0.0243; rbarD (adaptive) = 0.0746–0.2750], for P. grosskopfii and P. yuma (Table 2), suggesting that the signals of selective sweep are explained by a force acting differentially between stocks.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Linkage disequilibrium (LD) inferred through the standardized index of association (rbarD) for neutral and outlier (putative adaptive) loci datasets, under two hierarchical levels of genetic variation: including both differentiated genetic stocks in the analysis and within each stock or subpopulation.

https://doi.org/10.1371/journal.pone.0351301.t002

Genetic diversity and demographic inferences

The genetic diversity also revealed differences between the two stocks in the Pimelodus species. The observed heterozygosity for both species was always slightly lower in Stock 1 (0.28–0.31) than in stock 2 (0.44–0.52) (Table 3); being contrary to the behavior of expected heterozygosity, which was slightly similar for all stocks throughout the species (0.31–0.37) (Table 3). In relation to the inbreeding index (F_IS) for both species, this was between 8.7% − 9% in stock 1, while it showed an excess of heterozygotes, with negative values, for Stock 2 (Table 3). Regarding the effective population size (Ne), it could only be estimated in stock 1 of both species, due to the sensitivity of the algorithm with small sizes of the sample in which it is estimated, as the case of stock 2. Thus, in P. grosskopfii the Ne reached 379.3, while in P. yuma it was 313.8.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Diversity indexes and demographic parameters for genetic stocks of Pimelodus grosskopfii and Pimelodus yuma in the Magdalena-Cauca basin.

https://doi.org/10.1371/journal.pone.0351301.t003

After the migration model tests between stocks in Migrate-n, we found that in P. grosskopfii, the unidirectional pattern was the best model explaining migration between stocks (lmL = −36156,82), in this case, from stock 2 toward stock 1 (2Nm = 4.3 migrants per generation, S8 Table; Table 3); while the bidirectional model was the best one explaining migration between stocks in P. yuma (lmL = −23576,68), with 2.74 migrants per generation from the stock 1 toward stock 2, and 5.99 migrants per generation from stock 2 toward stock 1 (S8 Table; Table 3).

Likewise, the Bayesian skyline plot (BSP) (Figs 6 and 7) showed that the most recent common ancestor for stock 1 in both species was detected between 4.5–5.5 MYA (Late Miocene-Pliocene), suggesting an older origin for this stock. From there, in P. grosskopfii, stock 1 experienced a population expansion until ~3.5 MYA, when it stabilizes and remains constant until the present (Fig 6A). In the case of P. yuma, stock 1 did not experience expansion but remained constant until ~0.125 MYA (Fig 7A), from which time it suffers a reduction in population size until the present, which, even reduced, remained higher than that of stock 2. Particularly, in the case of stock 2 for both species, their most recent common ancestor was detected between ~1.7–2.5 MYA (Early Pleistocene). From ~1.2 MYA (Middle Pleistocene), in P. grosskopfii, stock 2 experienced a population expansion until ~0.24 MYA (Late Pleistocene), when it already stabilizes and remains constant until the present. In the case of P. yuma, stock 2 does not experienced expansion but remains constant from its detection until ~1.0 MYA (Middle Pleistocene), a time from which, unlike P. grosskopfii, it suffers an abrupt reduction in population size until 0.4 MYA, with a subsequent rapid expansion that goes up to ~0.24 MYA (Late Pleistocene), from which time it stabilizes until the present.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Bayesian Skyline Plot (BSP) for Pimelodus grosskopfii populations from the Magdalena–Cauca basin.

(A) BSP for Stock 1. (B) BSP for Stock 2. The Y-axis represents the effective population size scaled by the mutation rate (Ne × μ), while the X-axis represents time in millions of years before present (MYA). The shaded area represents the 95% highest posterior density interval.

https://doi.org/10.1371/journal.pone.0351301.g006

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Bayesian Skyline Plot (BSP) for Pimelodus yuma populations from the Magdalena–Cauca basin.

(A) BSP for Stock 1. (B) BSP for Stock 2. The Y-axis represents the effective population size scaled by the mutation rate (Ne × μ), while the X-axis represents time in millions of years before present (MYA). The shaded area represents the 95% highest posterior density interval.

https://doi.org/10.1371/journal.pone.0351301.g007

Putative identity of selective SNPs loci and enrichment analyses

After performing identity analyzing in BLAST (GenBank), only six of the 44 SNPs for P. grosskopfii [11338, 51741, 52924, 54464, 54596, 54676 (see S9 Table)] could be attributed to a genic or proteinic identity [E3 ubiquitin/ISG15 ligase TRIM25 (TRIM25), E3 ubiquitin-protein ligase (HERC2), ATPase Phospholipid Transporting 10D (ATP10D), RAS p21 protein activator 4 (RASA4), Neurexin 1 (NRXN1), and, GTPase activating protein (SH3 domain) binding protein 1 (G3BP1)], with 86–100% of identity (E-value<0.00001). In the case of P. yuma, of the 15 SNPs tested, only two [8894, 71235 (see S10 Table)] found a match with a genic region [SH3 domain-containing kinase-binding protein 1 (SH3KBP1), Ribosomal Oxygenase 2 (RIOX2)], with 92–93% of identity (E-value<0.00001). In addition, the PANTHER results showed three molecular functions where the P. grosskopfii putative genes could be involved in (ATP-dependent Activity, Binding and Transporter activity) (see S11A Fig), and four biological processes (Biological regulation, cellular process, localization, and metabolic process) (see S11B Fig). In the case of P. yuma, only one molecular function could be detected (Catalytic activity) (see S11C Fig), as well as only one biological process (Cellular process) (see S11D Fig).

Finally, functional annotations derived from the PANTHER Classification System, corroborated with classifications from the Gene Ontology (GO) database and supported by AI-assisted interpretation (GSEA) indicated that the identified genes are associated with several well-characterized physiological processes. Specifically, TRIM25-like proteins are involved in ubiquitination and antiviral immune responses; RASA4 participates in Ras-mediated signaling pathways; G3BP1 is associated with stress granule formation during cellular stress; NRXN1 is involved in synaptic communication and neuronal signaling; ATP10D plays a role in phospholipid transport and membrane dynamics; and HERC2 is associated with ubiquitination, DNA repair, and cellular stress responses.

In P. yuma, SH3KBP1 is involved in intracellular signaling and vesicular trafficking, while RIOX2 participates in epigenetic regulation through histone modification and cellular differentiation processes.

Population genetic structure and the role of selection and neutral variation

Our results indicate high genetic connectivity for Pimelodus yuma and P. grosskopfii populations, suggesting that physical barriers do not significantly restrict gene flow across their examined distribution. Thus, a simple model of genetic differentiation structured by geographic distance (isolation-by-distance) is rejected, despite the broad sampling range of more than 1,600 km for P. grosskopfii and 1,200 km for P. yuma. Even at this broader scale, our analysis corroborates previous findings [7,8], confirming the coexistence of two genetic stocks in sympatry. This two-stock structure persists despite the high gene flow observed in each stock across the sampled geographic regions, suggesting that differentiation is driven by non-spatial factors, such as ecological adaptation.

Our findings suggest that these stocks likely originated from ecological adaptations, with Stock 2 representing a strongly selected subset derived from a larger, more diverse ancestral population (Stock 1). This scenario aligns with hypotheses of divergence via “migratory timing” [16], where gene flow occurs only among individuals successfully reproducing after reaching spawning grounds, modulated by environmental cues [55]. The concept of temporal isolation as a driver of differentiation is not novel for this genus; it had been previously suggested for both P. yuma [7] and P. grosskopfii [8], linked to the basin’s distinct seasonal and hydrological periods. This hypothesis is supported by key biological observations. For instance, Pimelodus exhibits peaks of ichthyoplankton presence only during rainy seasons in the Magdalena River basin [15]. Furthermore, a greater number of P. grosskopfii individuals have been observed participating in the migration during the first of the two main hydrological periods in the Magdalena River [3]. However, conclusive genetic evidence for such temporal structuring remains to be demonstrated.

Evidence from allopatric and sympatric breeding populations shows that differences in migratory behavior often coincide with genetic divergence [16]. Mechanisms in several animals like selection against intermediate phenotypes can reinforce reproductive isolation [16]. In the case of Pimelodus, we hypothesize that extreme phenotypes—ecotypes with migration and spawning timing optimized for one flood pulse—achieve high fitness via synchronized reproduction and resource exploitation. However, individuals with intermediate traits probably suffer timing mismatches, reduced offspring survival, and a fitness valley, reinforcing divergence and reproductive isolation amid gene flow. Though not fully understood, behavioral divergence in migration likely plays a role in population differentiation. This has been demonstrated in Pimelodidae species, where migration and reproductive cycles influence gene flow, as observed in P. maculatus across Uruguay, Tibagi, and Tietê rivers [62–64], and in Pseudoplatystoma corruscans in the São Francisco and Paraguay basins [64].

Interestingly, despite no spatial structure, nearly discrete genetic stocks occupying the same river sections during different or overlapping migratory periods have been reported in P. reticulatum in the Paraná–Paraguay Basin, and in P. grosskopfii and P. yuma in the Cauca River [7,8]. These cases highlight the role of seasonal migration in population divergence and reproductive isolation [16,17]. Studies have linked such migratory divergence to speciation, often resulting in strong genetic differentiation—as seen in sharks, sea turtles, salmonids, and warblers [16].

A recent study on Cyphocharax magdalenae, a Characiform species co-occurring with Pimelodus, also revealed a seasonal population structure comprising two genetic stocks [20]. In that study, sampling was seasonally stratified, supporting the association of one stock with dry seasons (Jan–Mar, Jul–Aug) and the other with rainy seasons (Apr–Jun, Oct–Dec). These findings provide additional evidence for the presence of temporal genetic structuring in Neotropical rivers, further supporting the patterns of seasonal divergence observed in our current assessment.

Although additional seasonally targeted sampling is needed for P. grosskopfii and P. yuma to confirm seasonal stock association (e.g., rainy–dry or rainy–rainy cycles), our study clearly shows a selective contribution to stock divergence. Outlier loci showed fixed alternative alleles between stocks (≥90% in one and ≤10% in the other), with F_ST and PCA variance increasing nearly sevenfold when adaptive loci were included. Moreover, linkage disequilibrium rose when both neutral and adaptive markers were analyzed jointly. In addition, removing adaptive loci from the SNP matrix slightly reduced structure, and admixture was observed in some individuals, yet overall differentiation remained.

Across multiple fish systems, selection has emerged as a recurrent mechanism promoting reproductive isolation and even speciation, operating through adaptive divergence in ecological traits, life-history strategies, and mating-related phenotypes [65–67].

Phylogenetic analysis using neutral loci confirmed this structure as historical. Stock 2 formed a well-supported distinct monophyletic group, suggesting long-term fixation of alleles and an independent evolutionary trajectory. Interestingly, phylogenetic evidence suggested that Stock 2 likely emerged from a dispersal event originating in Stock 1 rather than from vicariance, possibly accelerated by past climatic shifts. According to Block & Levine (2021) [68], climate-driven environmental changes can promote the dispersion and establishment of genotypes better suited to emerging local conditions.

Chronology of the selection and the genetic structuring

It is evident that a selective force plays a central role in shaping the genetic structure of both Pimelodus yuma and P. grosskopfii. Available evidence indicates that this selective pressure has likely operated persistently over evolutionary timescales, progressively intensifying in concert with the evolving climatic and hydrological configuration of the Magdalena basin, resulting in increasingly restricted gene flow between individuals from the two stocks up to the present. This pattern is supported by multiple lines of neutral genetic variation analyses, including principal component analysis (PCA), pairwise F_ST estimates, migration rate inferences from MIGRATE, ancestry proportions from STRUCTURE (albeit partially), and historical phylogeographic reconstruction using RAxML. A key indicator of historical divergence is the time of origin of Stock 2, which appears to have emerged more recently than Stock 1, as confirmed by phylogenetic reconstruction. The most recent common ancestor (MRCA) of Stock 2 in both species is estimated to date back to ~2 million years ago (Middle Pleistocene), indicating a substantial evolutionary trajectory despite its younger age compared to Stock 1 (4.5–5.5 MYA).

Additionally, Bayesian Skyline Plot reconstructions revealed a recent demographic expansion in Stock 2 for both species, estimated between 0.25–0.5 MYA for P. yuma, and 0.25–1 MYA for P. grosskopfii. This pattern suggests a recent intensification of selective pressure—likely linked to evolving climatic and hydrological conditions in the Magdalena basin—favoring adaptive genotypes with enhanced resource exploitation and environmental performance [69]. Once adaptation occurs, fitness advantages may arise from changes in life-history traits that enhance population growth [69]. Functionally, these advantages may stem from increased or more efficient resource utilization, with direct consequences for demographic trends.

Interestingly, the timing of these expansions aligns with major climatic fluctuations during the Pleistocene [70], particularly those associated with abrupt reorganizations of atmospheric circulation driven by latitudinal shifts of the Intertropical Convergence Zone (ITCZ) [71]. During glacial periods, the ITCZ shifted equatorward, while in interglacial it moved northward in response to Northern Hemisphere warming [72].

Although the ITCZ is a global phenomenon, its effects on precipitation and drought patterns are modulated by regional atmospheric and ecological conditions. Equatorial rainfall is not governed solely by ITCZ convergence, but also by local factors such as topography-induced convection, proximity to oceans, atmospheric jets and waves, moisture recycling, and variability in land cover and albedo [73,74]. In South America, the Andean uplift and atmospheric CO₂ decline significantly influenced the positioning of the ITCZ, as well as the evolution of the South American monsoon system and rainforests [75].

Indeed, the estimated formation periods of Stock 1 (~4–5 MYA) and Stock 2 (~1.8–2.5 MYA) coincide with two major episodes in the final geomorphological reconfiguration of the Magdalena Basin, driven by the last phases of the Andean orogeny. The first event involved the separation of the Magdalena Basin from the Maracaibo Basin [≤4.5 MYA; [76]], and the second from the Orinoco Basin [~2.5 MYA; [77]], contributing to the long-term geographic and hydrological isolation of its ichthyofauna, and likely also to a new regional climate reorganization.

Currently, the Magdalena–Cauca basin exhibits remarkable morphological, ecological, and climatic heterogeneity, shaped by extensive floodplains, high Andean ranges, and localized climatic events, which result from interactions between orographic features and seasonal rainfall influenced by the Pacific Ocean and moisture inflows from the Amazon and Orinoco basins [78]. Critically, the biannual passage of the Intertropical Convergence Zone (ITCZ) generates two rainy seasons per year [78], leading to major flooding periods that drive dual cycles of biological productivity in floodplains through increased water and sediment input [79]. This seasonal pattern triggers two annual peaks of reproductive migration in potamodromous fishes such as Pimelodus [15].

Rainy seasons and the probably asynchronic migratory response modeling stocks in Pimelodus

We hypothesize that the two rainy seasons in the Magdalena–Cauca Basin are key drivers of the observed selection patterns, potentially promoting reproductive isolation via asynchronous migratory behavior in Stock 2 (respect to the Stock 1) of both Pimelodus species. While a deeper genomic investigation is needed to confirm this, several lines of evidence support it: (i) ichthyoplankton from both species peaks during the rainy seasons [15], indicating biannual reproduction; (ii) captive observations suggest different individuals spawn in each peak (Atencio-García, 2025, pers. comm.); and (iii) evidence of putative selection on SNPs located near genes potentially associated with physiological processes underlying reproductive migration, including environmental stress, neuro-sensory pathways, immune modulation, and metabolic regulation, as identified in this study and previously reported in other migratory fishes [80–82].

The two largest cis-Andean South American river basins harboring the greatest diversity of Pimelodus species—the Orinoco and Amazon—exhibit a unimodal flood-pulse regime associated with reproductive migration in Pimelodus species [83–85]. In contrast, the Magdalena–Cauca basin exhibits bimodal regime due to ITCZ dynamics, with a second rainy season and flood-pulse (Oct–Dec) aligning with South American cis-Andean basins, and a first rainy season (Apr–Jun) unique to this trans-Andean River system [3,14,15]. Given the cis-Andean origin of P. grosskopfii and P. yuma ancestors [28], adaptation to respond reproductively to this first flood peak may be a recent, trans-Andean region-specific event. The emergence of Stock 2 likely reflects an evolutionary response to this novel environmental cue, as suggested by phylogenetic and demographic reconstructions showing a younger, dispersed derived lineage from Stock 1—the more diverse and ancestral population. It is known that dispersal evolution and local adaptation are eventually linked, affecting the range dynamics of species facing environmental changes associated with climate [68]. Genetic diversity and demographic analyses reveal lower heterozygosity and effective population size in Stock 2, consistent with strong selection effects [86,87]. Similar signatures of selection-driven divergence have been reported in Oncorhynchus nerka and Salmo trutta [88].

We hypothesize that this kind of striking selective variation in migration patterns probably modelling structure in Pimelodus, is a well-known mechanism of ecotype differentiation called “migration timing” [89–92] and could have a genetic base behind. For instance, in species like Chinook salmon, early and late migrants show distinct phenotypes (ecotypes) linked to regions near genes like ROCK1 and GREB1L [17]. While in salmonids a few genes explain most of the variation (85%) [88], in Pimelodus, adaptive divergence appears polygenic, with loci under selection accounting for ~70% of the variance, through an apparent selective sweep mechanism, as suggested by the LD test (rbarD) by using neutral and selective loci.

Interestingly, in this study, although gene flow exists, population structure remains. In our dataset, only ~15% of selective loci showed fixation patterns expected under Isolation by Adaptation (IBA) [65], while 67% showed high heterozygosity in one stock and homozygosity in the other—suggesting balancing selection or ongoing gene flow (See S12 and S13 Figs for more details). This aligns with the “isolation with migration” model [93,94], where likely the selection and drift act within populations, but gene flow counteracts divergence. A similar pattern was observed in Chinook salmon, where lineages or populations appear to be more genetically distinct than expected, despite clear evidence of interbreeding among them [92].

STRUCTURE analyses reveal that P. grosskopfii’s Stock 2 is highly differentiated (with <15% shared ancestry with Stock 1), likely due to restricted, unidirectional migration, while P. yuma’s Stock 2 (~50% shared ancestry) reflects more bidirectional gene flow. This differentiation may reflect either stronger selection in P. grosskopfii or a longer divergence time (~1.7 MYA) compared to P. yuma (~1 MYA), consistent with greater neutral allele drift over time [65] in P. grosskopfii. In addition, this pattern may also be explained by the retention of ancestral polymorphisms, whereby shared genetic variation inherited from a common ancestor can persist across populations after divergence, generating signals of coancestry even in the absence of ongoing gene flow [86,93]. As despite observable coancestry in P. yuma between the putative stocks, phylogenomic inferences support a historical differentiation between them.

Lastly, sampling year was excluded as a confounding structuring factor, given that individuals from both genetic stocks were often collected concurrently in the same sampling year, as shown in S1 Table.

Implications for conservation

Based on our findings, genetic Stocks 1 and 2 within each species should be regarded as both Management Units (MUs) and Adaptive Units (AUs), considering the consistent phylogenetic patterns and population structure derived from both neutral and adaptive loci [22,95]. Each stock represents a distinct lineage with ongoing gene flow or demographic connectivity.

Stock 1, in both species, shows higher abundance, older evolutionary persistence, and more stable diversity indices, as well as moderate effective population size (Ne = 50–500) [96], indicating a reduced risk of severe inbreeding effects in the short term. It is thus the more resilient and genetically healthy stock. In contrast, Stock 2 appears younger, less frequent in the basin, and under stronger selective pressure. Although its Ne could not be directly estimated, Bayesian Skyline Plots suggest a smaller effective size relative to Stock 1. These features suggest Stock 2 is at higher risk of inbreeding depression and reduced reproductive viability (Ne < 50–100), making it a conservation priority [96–98].

Nonetheless, neither stock meets the minimum Ne threshold (>500) required to ensure long-term adaptive potential [96,98], meaning both are below the level needed to maintain evolutionary resilience. These observations warrant caution, and further studies with increased sample sizes are recommended to confirm the current patterns.

Additionally, diversity indices were lower compared to previous studies using microsatellites in the same basin for both species [7,8]. For instances, in P. grosskopfii, observed heterozygosity was 43.4% lower and the inbreeding coefficient was 44.8% lower than estimates obtained using microsatellites. Similarly, P. yuma showed reductions of 42.5% in heterozygosity and 41.0% in the inbreeding coefficient relative to microsatellite-based estimates. These differences reflect the biallelic nature of SNPs, which typically underestimate genetic diversity compared to multiallelic microsatellites [99]. Therefore, a correction factor of approximately 41–45% may be considered when comparing diversity estimates between marker types, particularly when interpreting SNP-based results regarding these species. Nevertheless, SNPs offer higher genomic resolution and reduced bias in estimating population parameters by capturing both neutral and adaptive processes, making them a powerful complement to SSRs in conservation genetics [95].