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 9 / 14 Visual diagnostics reveal latent structure While numerical metrics consistently suggested low clustering validity under realistic scenarios, dimension reduction tech- niques (PCA and UMAP) revealed subtle groupings in certain algorithms (Fig 3 and S1 File). Notably, DBSCAN and Affin- ity Propagation identified structure congruent with known groupings, despite low ARI/NMI scores. PCA explained >80% of variance on the first two axes, underscoring biologically relevant structure that escapes conventional metrics. This discor- dance highlights the limitations of relying solely on statistical indices when interpreting weakly differentiated datasets. Discussion Our study presents a rigorous and biologically informed evaluation of clustering methods applied to phenotypic datasets from neglected and underutilized species (NUS), exemplified by fonio. It reveals several critical insights: 1. Fundamental methodological mismatch: Under realistic phenotypic simulations (Pst ≈ 0.16; group sizes 91, 43, 46), all tested clustering algorithms—ranging from partition-based (K-means, PAM) to model-based (GMM) and density-based methods (DBSCAN, HDBSCAN)—yielded uniformly poor agreement with true group labels (ARI < 0.06; Fig 3. PCA and UMAP reveal latent structure beyond numerical metrics. Visual representation of clustering patterns across dimensionality reduc- tion methods, indicating underlying structure undetected by ARI and NMI. https://doi.org/10.1371/journal.pone.0329254.g003