Relative performance of cluster algorithms and validation indices in maize genome-wide structure patterns

A number of clustering algorithms are available to depict population genetic structure (PGS) with genomic data; however, there is no consensus on which methods are the best performing ones. We conducted a simulation study of three PGS scenarios with subpopulations k = 2, 5 and 10, recreating several maize genomes as a model to: (1) compare three well-known clustering methods: UPGMA, k-means and, Bayesian method (BM); (2) asses four internal validation indices: CH, Connectivity, Dunn and Silhouette, to determine the reliable number of groups defining a PGS; and (3) estimate the misclassification rate for each validation index. Moreover, a publicly available maize dataset was used to illustrate the outcomes of our simulation. BM was the best method to classify individuals in all tested scenarios, without assignment errors. Conversely, UPGMA was the method with the highest misclassification rate. In scenarios with 5 and 10 subpopulations, CH and Connectivity indices had the maximum underestimation of group number for all cluster algorithms. Dunn and Silhouette indices showed the best performance with BM. Nevertheless, since Silhouette measures the degree of confidence in cluster assignment, and BM measures the probability of cluster membership, these results should be considered with caution. In this study we found that BM showed to be efficient to depict the PGS in both simulated and real maize datasets. This study offers a robust alternative to unveil the existing PGS, thereby facilitating population studies and breeding strategies in maize programs. Moreover, the present findings may have implications for other crop species.