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 6 / 14 Key parameter choices included: K-means (nstart = 25), DBSCAN/HDBSCAN (minPts = 5 with a data-driven eps heuris- tic), Fuzzy C-means (m = 2), Two-Step clustering (α = 0.05 trimming), and SOM (rlen = 100 learning steps). No dimensionality reduction (e.g., PCA) was applied prior to clustering, as this would artificially remove the empirically calibrated inter-trait correlations that characterize real phenotypic datasets of neglected and underutilized species (NUS). The framework intentionally preserves these biological dependencies to maintain ecological realism rather than optimizing for computational convenience. A comprehensive summary of all algorithmic settings—including package versions, R functions, and parameter specifi- cations—is provided in S1 File (Supplementary Information). All implementations used a unified wrapper function (apply_ clustering()), and convergence followed the default internal criteria of the respective R functions. Evaluation metrics Clustering performance was evaluated using four complementary indices. • Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI) quantify external agreement between pre- dicted and true population labels, corrected for chance [12,15]. • Silhouette Coefficient [26] and Davies–Bouldin Index [27] provide internal, label-independent assessments of cohe- sion and separation, enabling evaluation under weak or diffuse population structure.

Together, these external and internal metrics capture both supervised and unsupervised aspects of clustering quality, bridging algorithmic and biological interpretability. A dispersion statistic (betadisper; [28]) was also computed to quantify within-group variability relative to population centroids. Statistical analysis and reproducibility Each simulation scenario was replicated 100 times. Performance metrics were summarized using means and standard deviations; robustness was quantified as the inverse of inter-replicate variance. Pairwise comparisons employed Dunn’s tests with FDR correction (α = 0.01). Visualization of clustering validity employed Principal Component Analysis (PCA) and Uniform Manifold Approximation and Projection (UMAP) [29,30]. UMAP projections were computed using the umap() function (R package umap, v0.2.10.0) with parameters n_neigh- bors = 15, min_dist = 0.1, metric = “euclidean”, and a fixed random seed for reproducibility. These settings were retained because they offer a biologically sensible balance between preserving global structure and revealing local neighborhoods in medium-sized phenotypic datasets (~100–300 accessions). In UMAP, n_neighbors determines the balance between local versus global manifold structure; the official documentation recommends the default value of 15 as a reasonable compromise for general-purpose use, while min_dist = 0.1 prevents excessive crowding and preserves continuous gradi- ents typical of quantitative trait data.

These defaults are widely adopted in biological applications, and recommended practice in single-cell and phenotypic visualization places n_neighbors at approximately 3–10% of the sample size—corresponding closely to 15 neighbors for n ≈ 180. Accordingly, our parameterization aligns with established empirical practice and ensures stable, interpretable embeddings for phenotypic datasets of neglected crops [31–33]. All analyses were performed in R (version 4.3.0). Details of the statistical procedures, algorithms, and evaluation met- rics are described above. Results Clustering performance under realistic phenotypic conditions Under phenotypic conditions empirically calibrated from West African fonio landraces (Pst ≈ 0.16), all eleven tested clustering algorithms exhibited uniformly low performance across validation metrics (Table 1 and Fig 1). The Adjusted