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 3 / 14 settings. This structural misalignment directly undermines biodiversity conservation targets. As demonstrated by Arneth et al. [20], climate pressures threaten to invalidate preservation strategies for NUS based on artificial phenotypic clusters, even when controlling for non-climatic pressures like habitat exploitation and ecological fragmentation. To address this methodological gap, we introduce a biologically grounded simulation framework tailored to the statistical and ecological realities of neglected and underutilized species (NUS). Calibrated with empirical phenotypic data from West African fonio (Digitaria exilis), the framework enables rigorous benchmarking of eleven clustering algo- rithms—including K-means, PAM, Ward’s method, GMM, DBSCAN, HDBSCAN, Spectral clustering, Fuzzy C-means, Affinity Propagation, Self-Organizing Maps (SOM), and Two-Step clustering—under both idealized and realistic con- ditions. The simulated datasets reproduce key real-world complexities: heterogeneous trait distributions (normal and gamma), strong inter-trait correlations (up to r = –0.84), heteroscedastic variances, and moderate population structure (Pst ≈ 0.15). We evaluate algorithm performance across 100 simulation replicates using ARI, NMI, Silhouette coefficient, and Davies–Bouldin index, supplemented by dimensionality reduction tools such as PCA and UMAP to assess the geo- metric structure of clusters.

Our central hypothesis is that apparent clusters in phenotypic datasets of NUS with modest genetic structure often arise as statistical artefacts driven by methodological constraints, rather than reflecting genuine biological partitions. By testing this systematically, our study offers a reproducible, R-based diagnostic tool to evaluate clustering validity before drawing evolutionary or agronomic conclusions. We advocate for a paradigm shift: empirically calibrated simulations should replace idealized Gaussian models as the standard for evaluating clustering performance in NUS research. This shift prioritizes biological realism over compu- tational convenience, particularly in regions such as sub-Saharan Africa, where genotyping is often inaccessible, while phenotyping remains more feasible and cost-effective. The tool is open-source, modular, and designed for reproducibility and cross-species adaptability. To maximize generalizability, its parameters were not only calibrated from fonio, but also designed to accommodate empirical data from other NUS available in the literature. This article thus serves as the methodological cornerstone of a broader research agenda aimed at systematically benchmarking clustering approaches for biodiversity assessment across diverse underutilized taxa. It promotes a phenotype-first paradigm in agrobiodiversity research—one that centers ecological realism, supports data-driven decision- making, and advances conservation and breeding efforts in resource-limited settings. Although calibrated using fonio as

a model, the framework is crop-agnostic and can be readily reparameterized for other neglected species by adjusting the trait covariance structure, differentiation level (Pst), and distributional assumptions. This flexibility makes it suitable as a general diagnostic platform for phenotypic clustering in data-scarce crops. Materials and methods Empirical data and trait architecture This study draws upon agromorphological trait data from 180 Digitaria exilis (fonio) landraces characterized by Bio et al. [1], spanning three geographically distinct groups (n = 91, 43, and 46). Eight traits were selected for their agronomic relevance and quantitative tractability: plant height (PHT), number of internodes (NIN), flowering time (FLO), maturity time (MAT), number of grains per inflorescence (NGR), grain yield per plant (GRY), harvest index (HI), and thousand-seed weight (TSW). All empirical parameters were manually extracted from the published tables of Bio et al. [1], Journal of Crop Improvement 34(4):512–530). Specifically, group sizes, trait means, standard deviations, and the inter-trait correlation matrix were taken directly from the reported agromorphological dataset. These empirical statistics were used to recon- struct the phenotypic covariance structure and served as calibration inputs for the simulation of realistic fonio populations. Trait distributions were predominantly normal, with the exception of GRY, which followed a gamma distribution. Empirical trait correlations (range: r = –0.84 to 0.77) were encoded in a near-positive definite correlation matrix [21], and

group-specific variances were retained to reproduce heteroscedastic structures in simulations.