Empirically calibrated simulations reveal the limits of phenotypic clustering algorithms for biodiversity assessment in data-scarce crops

PLOS One | https://doi.org/10.1371/journal.pone.0329254 December 17, 2025 4 / 14 Simulation framework Phenotypic differentiation (Pst) estimation and calibration. Phenotypic differentiation (Pst ) was computed as the proportion of total phenotypic variance explained by between-group differences, analogous to Wright’s F st for quantitative traits [22]. The phenotypic covariance matrices used for Pst calibration were derived from the empirical correlation structure of Bio et al. [1], ensuring that simulated trait dependencies reproduce the inter-trait architecture observed in real fonio landraces. For a trait matrix X with group labels g, Pst was calculated as: Pst = VB VT where VB is the trace of the between-group covariance matrix (inter-group variance component) and VT is the trace of the total covariance matrix. In simulations, the between-group covariance matrix was iteratively adjusted using a scaling factor until the realized Pst converged to the target value—0.15 for the realistic scenario and 0.80 for the idealized scenario. Convergence was accepted when | Prealized st – Ptarget st |< 0.01. Across 100 replicates, the mean realized Pst was 0.16 ± 0.001 for the realistic scenario, confirming accurate calibration. This procedure ensures that the simulated datasets reproduce the moderate phenotypic differentiation observed in fonio landraces [1], while maintaining the framework’s flexibility for other neglected crops by allowing adjustment of the target Pst parameter. The calibration procedure is implemented in the function simulate_populations(), which computes Pst

through the helper function calc_fst_pheno(). Rationale for the Pst target and comparability with Fst. The realistic-scenario target of Pst ≈ 0.15 was chosen to represent a biologically plausible level of moderate differentiation commonly observed in smallholder and landrace systems, where recurrent seed exchange, open pollination, and overlapping agroecological conditions limit genetic divergence. Conceptually, Pst—the proportion of phenotypic variance attributable to between-group differences—is an observable proxy for Qst when heritabilities are unknown. However, comparisons to neutral-marker Fst must be interpreted with caution, because Pst also depends on environmental variance and the ratio of additive genetic variances within versus among groups [23,24]. Our choice of approximately 0.15 thus reflects a heuristic midpoint rather than a strict numerical equivalence to genetic differentiation, but it aligns closely with the low-to-moderate genetic structure recently quantified in fonio. In a genomic study of Digitaria exilis populations across Senegal, Diop et al. [25] reported a mean Fst = 0.054 (range −0.038 to 0.18), indicating weak but significant differentiation among ethnolinguistic groups. This pattern reflects shared evolutionary and management processes—continuous gene flow, recurrent seed exchange, and stabilizing selection under comparable ecological conditions—typical of farmer-managed landrace systems in West Africa. Importantly, Diop et al. [25] demonstrated that social and cultural boundaries (ethnic and linguistic groups) explained

more of the genetic structure than geography or environment, and that the highly self-pollinating mating system of fonio (observed heterozygosity ≈ 0.05) contributes to maintaining intra-group uniformity while preserving subtle inter-group diver- gence. These findings reinforce the biological realism of our simulation framework, in which moderate phenotypic differen- tiation arises from socially mediated but genetically connected populations. We therefore explicitly avoid relying on earlier studies that reported inflated differentiation values based on dominant markers (e.g., AFLP), which are unsuitable for calibrating a low-structure regime. Instead, our empirical grounding in Diop et al. [25] ensures that the simulated scenarios faithfully represent the weakly structured, socially stratified population architecture characteristic of neglected and underutilized crops. Two simulation scenarios were implemented to assess clustering robustness: