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 5 / 14 1. Realistic scenario: Phenotypic datasets (n = 180) were synthetically generated using a multivariate transformation pipeline: • Correlated base generation: Traits simulated with MASS::mvrnorm() to match empirical covariance structure. • Distributional alignment: Gamma-distributed traits (e.g., GRY) transformed from normal deviates via quantile matching. • Population structure: Group-level covariance matrices were iteratively tuned to achieve a target phenotypic differentia- tion of Pst ≈ 0.15 (mean realized: 0.16 ± 0.001). • Heteroscedasticity: Trait-wise standard deviations were scaled per group using empirical multipliers (e.g., Pop1 GRY SD × 0.8). 2. Idealized control: A benchmark dataset featuring: • Balanced group sizes (n = 50 per group), • High between-group differentiation (Pst = 0.80), • Independent traits (identity matrix), • Homoscedastic variances. Clustering algorithms We benchmarked eleven clustering methods representing major analytical paradigms: • Partitioning: K-means, PAM; • Hierarchical: Ward’s method; • Model-based: Gaussian Mixture Models (GMM); • Density-based: DBSCAN, HDBSCAN; • Graph-theoretic: Spectral clustering; • Affinity-based: Affinity Propagation; • Fuzzy clustering: Fuzzy C-means; • Neural-network-based: Self-Organizing Maps (SOM); • Hybrid scalable: Two-Step clustering. For parametric methods, the number of clusters (k = 3) was fixed based on known population structure [1]. Non-parametric

methods relied on data-driven heuristics. Parameter tuning followed best-practice guidelines (e.g., DBSCAN’s eps set to 90th percentile of kNN distances; minPts = 5). Algorithmic settings and reproducibility All clustering methods were applied to z-score standardized trait data using identical preprocessing and replication schemes. For algorithms requiring predefined cluster numbers, k = 3 was fixed based on the known population structure of fonio landraces [1]. Random seeds were controlled throughout the simulation pipeline to ensure full reproducibility.