Abstract
BACKGROUND: Whole-genome sequencing (WGS) data contain a large proportion of genetic information for farm animals, while machine learning algorithms offer a powerful approach for discovering population-informative genetic markers for population assignment. These methods can enhance population classification, supporting genetic resource conservation and optimizing breeding strategies. METHODS: WGS data from 187 individuals representing eight water buffalo breeds (four river type breeds, i.e., Mediterranean, Murrah, Nili Ravi, and Binglangjiang, and four swamp type breeds, i.e., Diandongnan, Dehong, Guizhoubai, and Shanghai) were analyzed. Multiple single nucleotide polymorphism (SNP) selection methods including the population genetic differentiation index (F(ST)), the machine learning-based Minimum Redundancy Maximum Relevance (mRMR) and Relief-F algorithm were investigated to pinpoint the highly informative SNPs for population assignment. Four machine learning classifiers-Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Naive Bayes (NB), and Random Forest (RF)-were applied for population assignment of the eight water buffalo breeds. Nine different SNP set sizes (100, 250, 500, 800, 1000, 2000, 3000, 5000, 10000) were examined. The impact of SNP selection methods, classifiers and SNP set size on assignment accuracy was evaluated. RESULTS: Population genetic analysis revealed hierarchical structures among the eight water buffalo breeds, with two divergent groups (swamp type buffaloes and river type buffaloes) and relatively closely related breeds within each group. Among the classifiers, SVM and NB outperformed RF and KNN in terms of assignment accuracy. Accuracy improved with an increasing number of SNPs, reaching over 85% for all classifiers with 10,000 SNPs. The mRMR method provided the highest accuracy, achieving 97.6% with 2000 SNPs when paired with the SVM classifier, followed by Relief-F (94%) and F(ST) (90.4%). A panel of 768 mRMR-selected SNPs achieved 98.8% assignment accuracy using SVM classifiers, indicating it is a robust tool for reliable population assignment. CONCLUSION: By integrating machine learning-based feature selection with machine learning classifier, this study provides an effective framework for mining highly informative genetic markers for population assignment from whole-genome sequencing data.