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 12 / 14 For methodological development: • Establish detectability thresholds. Use the simulation platform to explore how clustering performance changes with increasing phenotypic differentiation (Pst), identifying thresholds where structure becomes reliably detectable. Such detectability curves can guide algorithm choice and interpretation in other neglected crops. We encourage practitioners to apply this diagnostic framework before making conservation or breeding decisions based on phenotypic clustering—particularly for NUS where genomic validation may be unavailable. This approach bridges methodological rigor and applied relevance, strengthening the reliability of phenotypic inferences in biodiversity research. Conclusion This study demonstrates that conventional clustering algorithms systematically underperform when applied to realistically simulated phenotypic data from neglected and underutilized species (NUS), such as fonio. These results reveal not only methodological limitations, but also a deeper misalignment between algorithmic assumptions and the biological realities of NUS datasets—characterized by overlapping traits, heteroscedasticity, and modest differentiation. Critically, our analyses show that low ARI and NMI scores can obscure meaningful biological structure, which remains detectable through visual and geometric diagnostics such as PCA and UMAP. This highlights the need to treat clustering

as an exploratory rather than confirmatory tool, combining quantitative indices with qualitative and biological interpretation. The biologically calibrated simulation framework developed here offers a reproducible, flexible diagnostic platform to benchmark clustering performance under realistic conditions. It enables researchers to align methodological choices with empirical data properties, fostering more robust and interpretable phenotypic stratification. Future research should focus on developing biologically grounded validation metrics better suited to continuous and weakly structured biological data, and on systematically mapping clustering performance across gradients of phenotypic differentiation (Pst). Integrating missing-data mechanisms (MCAR/MAR) and genomic validation will further strengthen the framework’s applicability to real-world datasets. Ultimately, this simulation-based diagnostic platform is crop-agnostic and extensible across species, providing a foun- dation for data-driven standards in phenotypic structure detection. By bridging methodological rigor with biological realism, it offers a path toward more transparent, reproducible, and evolutionarily meaningful biodiversity assessment in data- scarce crops. Supporting information S1 File. Supplementary figures and table. This compressed archive contains Figures S1–S12 and Table S1. (ZIP) Acknowledgments The author expresses sincere appreciation to Dr. Devra I. Jarvis and the Raffaella Foundation for their continuous efforts

to promote research and capacity building on agrobiodiversity in West Africa and beyond. The present work aligns with this broader vision, although it was conducted independently within the author’s academic affiliation at the University Abdou Moumouni. Author contributions Conceptualization: Abdel Kader Naino Jika. Data curation: Abdel Kader Naino Jika. Formal analysis: Abdel Kader Naino Jika.