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 7 / 14 Rand Index (ARI) ranged from –0.006 (HDBSCAN) to 0.053 (GMM), indicating near-random agreement with true groupings. Negative ARI values indicate that agreement between predicted and true partitions is slightly lower than would be expected by chance. This occurs when clusters are statistically misaligned under weak phenotypic differ- entiation, and should not be interpreted as algorithmic failure but rather a signal of diffuse or overlapping biological structure. Normalized Mutual Information (NMI) values were similarly low (maximum: 0.083, Affinity Propagation), suggesting poor information retention. Silhouette coefficients exceeded the interpretability threshold (0.25) in only one case (DBSCAN: 0.379), while Davies–Bouldin indices consistently exceeded 1.2, signifying poor inter-cluster separation. Importantly, these results were consistent across 100 simulation replicates per algorithm (except HDBSCAN: n = 89 due to convergence issues), reinforcing the reliability of observed patterns. Despite modest differences between methods, no algorithm achieved a biologically meaningful clustering outcome under realistic trait complexity and moderate population differentiation. Table 1. Mean clustering metrics under realistic conditions (± SD; n = 100 unless noted). Méthode ARI NMI Silhouette Davies-Bouldin GMM 0.053 ± 0.064 0.070 ± 0.062 0.127 ± 0.059 1.768 ± 0.516 TwoStep 0.051 ± 0.058 0.063 ± 0.048 0.188 ± 0.032 3.641 ± 1.347

Spectral 0.045 ± 0.066 0.055 ± 0.052 0.240 ± 0.057 1.429 ± 0.862 PAM 0.045 ± 0.050 0.058 ± 0.050 0.241 ± 0.025 1.289 ± 0.090 Fuzzy C-means 0.042 ± 0.047 0.055 ± 0.048 0.244 ± 0.023 1.288 ± 0.083 SOM 0.042 ± 0.045 0.057 ± 0.048 0.262 ± 0.022 1.223 ± 0.079 K-means 0.041 ± 0.044 0.056 ± 0.047 0.263 ± 0.021 1.223 ± 0.077 Ward.D2 0.040 ± 0.044 0.056 ± 0.045 0.243 ± 0.029 1.249 ± 0.124 Affinity Propagation 0.031 ± 0.019 0.083 ± 0.028 0.132 ± 0.011 1.556 ± 0.087 DBSCAN 0.007 ± 0.011 0.009 ± 0.008 0.379 ± 0.073 3.184 ± 2.226 HDBSCAN (n = 89) −0.006 ± 0.031 0.033 ± 0.026 −0.153 ± 0.132 4.648 ± 2.169 https://doi.org/10.1371/journal.pone.0329254.t001 Fig 1. Mean performance (± SD) of eleven clustering methods across ARI, NMI, Silhouette, and Davies–Bouldin indices in simulations cali- brated with empirical fonio trait data (Pst ≈ 0.16). https://doi.org/10.1371/journal.pone.0329254.g001