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 8 / 14 Performance under idealized conditions In the idealized scenario featuring high differentiation (Pst ≈ 0.80), independent traits, equal sample sizes, and homosce- dasticity, algorithm performance dramatically improved (Table 2 and Fig 2). Parametric methods such as GMM, Fuzzy C-means, and K-means achieved near-perfect ARI (>0.97) and NMI (>0.96) values, confirming both the integrity of the simulation pipeline and the sensitivity of these algorithms to favorable statistical conditions. Silhouette coefficients (~0.39) and Davies–Bouldin indices (~1.05) indicated well-separated, coherent clusters. By contrast, non-parametric methods (e.g., DBSCAN, HDBSCAN, Affinity Propagation) showed lower mean perfor- mance and higher variance, suggesting reduced reliability even under ideal conditions. DBSCAN failed to converge in 4% of replicates due to poor density parameter fits. Table 2. Mean clustering metrics under ideal conditions (± SD; n = 100). Méthode ARI NMI Silhouette Davies-Bouldin GMM 0.986 ± 0.026 0.981 ± 0.033 0.394 ± 0.077 1.051 ± 0.219 Fuzzy C-means 0.976 ± 0.042 0.968 ± 0.052 0.394 ± 0.076 1.048 ± 0.215 K-means 0.975 ± 0.045 0.968 ± 0.053 0.394 ± 0.076 1.047 ± 0.213 Ward.D2 0.955 ± 0.076 0.949 ± 0.077 0.390 ± 0.080 1.061 ± 0.234 PAM 0.948 ± 0.073 0.941 ± 0.077 0.390 ± 0.080 1.058 ± 0.227 Spectral 0.937 ± 0.134 0.936 ± 0.119 0.372 ± 0.101 1.102 ± 0.293 SOM 0.875 ± 0.212 0.894 ± 0.164 0.371 ± 0.086 1.142 ± 0.292

TwoStep 0.862 ± 0.119 0.790 ± 0.093 0.340 ± 0.079 2.339 ± 0.478 HDBSCAN 0.652 ± 0.276 0.663 ± 0.210 0.292 ± 0.147 2.317 ± 1.223 DBSCAN 0.436 ± 0.391 0.437 ± 0.378 0.282 ± 0.122 1.775 ± 1.011 Affinity Propagation 0.415 ± 0.101 0.496 ± 0.069 0.136 ± 0.027 1.856 ± 0.174 https://doi.org/10.1371/journal.pone.0329254.t002 Fig 2. Validation of simulation pipeline and algorithmic capacity under favorable statistical assumptions: independent traits, strong structure, and homoscedasticity. https://doi.org/10.1371/journal.pone.0329254.g002