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 10 / 14 NMI < 0.09). This quantitatively demonstrates the fundamental limitations noted by Jombart et al. [6] and Odong et al. [7], extending their qualitative insights to simulation-based rigor. The small negative ARI values observed for some algorithms further confirm that, under these weakly structured conditions, agreement with the true partition can fall below what would be expected by chance alone. The reliance on assumptions of discrete, isotropic, homogeneous clusters in these algorithms is misaligned with the overlapping, correlated, heteroscedastic, and weakly differentiated structures typical of NUS phenotypic data [1,14]. 2. ARI/NMI fail to detect subtle biological structure: Despite consistently low ARI and NMI scores, PCA and UMAP visualizations revealed coherent group separation in some clustering outputs—capturing over 80% of variance on the first two axes. This uncoupling reinforces critiques by Romano et al. [13] and Vinh et al. [12], showing that conventional metrics inherently underestimate biologically meaningful differentiation when clusters are continuous or overlapping. We argue these low scores should be viewed as signals of underlying biological complexity, rather than algorithmic failure. 3. Visualization alone is insufficient: Diagnostics must combine metrics and plots: Our findings underscore the necessity of integrating qualitative (PCA/UMAP) and quantitative diagnostics. Despite pronounced visual structure in

DBSCAN and Affinity Propagation outputs, reliance solely on ARI/NMI would reject these methods. This aligns with cri- tiques in single-cell genomics and suggests biodiversity analyses require multidimensional validation protocols [30,29]. 4. Impact of PCA preprocessing on clustering performance: A common practice in applied phenotypic analyses is to perform clustering on principal component scores rather than raw traits. In our simulations, this would artificially remove the inter-trait correlations that were deliberately modeled from empirical fonio data to represent real biological dependencies. While PCA-based preprocessing may improve numerical stability or visual interpretability, it would not resolve the fundamental issue of weak phenotypic differentiation (Pst ≈ 0.15). Therefore, we expect that clustering on PCA- reduced data would yield marginally higher scores but would not alter the overall conclusion that conventional algorithms fail under biologically realistic conditions. 5. Trade-off between stability and accuracy: Analyzing robustness (inverse ARI variance) exposes a critical practical trade-off: high-performing but unstable methods (e.g., GMM) versus more stable, interpretable algorithms (e.g., PAM, Ward’s hierarchical). This finding supports prioritizing reproducibility and interpretability in applied NUS research—a stance aligned with Pilling et al. [5] and Vodouhè et al. [4]. 6. Toward data-driven clustering via empirical simulation: Crucially, our framework demonstrates the value of

context-specific simulation. By parameterizing distributions (normal, gamma, binomial), sample imbalance, correlation structures, heteroscedasticity, and group number, researchers can identify clustering methods most robust to specific dataset properties and assumption violations. This moves phenotypic clustering from a one-size-fits-all approach to a tailored, data-driven strategy. 7. Need for novel clustering validation metrics: The observed mismatch between conventional validation metrics and visual/biological structure highlights the urgent necessity for new indices—designed for weakly structured NUS data. Such metrics must integrate geometric separation, internal cohesion, and be robust to non-normality, unbalanced designs, heteroscedasticity, and nonlinear correlations. 8. Mapping performance across population differentiation dradient: Future work should focus on how clustering detectability varies with phenotypic differentiation. As suggested by Jombart et al. [6], critical Pst thresholds may exist below which conventional clustering algorithms become unreliable. The present study concentrates on a single empir- ically grounded differentiation level (Pst ≈ 0.15), but future analyses should systematically map clustering performance across a continuum of differentiation values.