Abstract
As the scale of deep whole-genome sequencing (WGS) data has grown exponentially, hundreds of millions of single nucleotide polymorphisms (SNPs) have been identified in livestock. Utilizing these massive SNP data in population stratification analysis, ancestry prediction, and breed diversity assessments leads to overfitting issues in computational models and creates computational bottlenecks. Therefore, selecting genetic variants that express high amounts of information for use in population diversity studies and ancestry inference becomes critically important. Here, we develop a method, HITSNP, that combines feature selection and machine learning algorithms to select high-representative SNPs that can effectively estimate breed diversity and infer ancestry. HITSNP outperforms existing feature selection methods in estimating accuracy and computational stability. Furthermore, HITSNP offers a new algorithm to predict the number and composition of ancestral populations using a small number of SNPs, and avoiding calculating the number of clusters. Taken together, HITSNP facilitates the research of population structure, animal breeding, and animal resource protection.