On Assigning Individuals from Cryptic Population Structures to Optimal Predicted Subpopulations: An Empirical Evaluation of Non-parametric Population Structure Analysis Techniques

Abstract

Many algorithms have been proposed to analyze population structures from the single nucleotide polymorphism (SNP) genotyping data of some number of individuals and try to assign individuals to genetically similar groups. These algorithms can be categorized into two computational paradigms: parametric and non-parametric approaches. Although the parametric-based approach is a gold standard for population structure analysis, the computational burden incurred by running these algorithms is unacceptable for large complex dataset. As genotyping platforms incorporating more SNPs, analyzing ever larger and more complex datasets are becoming a standard practice. Hence, the computationally efficient non-parametric methods for analysis of genotypic datasets are needed to reveal the population structure. In this study, we evaluated two leading non-parametric population structure analysis techniques, namely ipPCA and AWclust, on their abilities to characterize the genetic diversity and population structure of two complex SNP genotype datasets (as many as 243855 SNPs). The head-to-head comparisons were conducted on two major aspects: ability to infer the number of genetically related subpopulations (K) and ability to correctly assign individuals to these subpopulations. The experimental results suggested that AWclust could be more suitable when applying to a small and less complex dataset. However, with a large and more complex dataset, ipPCA is a much better choice yielding higher accuracy on assigning genetically similar individuals to the inferred groups.