Machine learning-assisted genomic characterization of population differentiation between Chinese Kirghiz and Kyrgyzstan local sheep

Chinese Kirghiz (KIR) sheep inhabit the eastern Pamir region near the China–Kyrgyzstan border, but their genomic relationship with geographically adjacent sheep populations in Kyrgyzstan remains incompletely resolved. We characterized population genomic differentiation between KIR and Kyrgyzstan local sheep while explicitly evaluating the effects of sample-size imbalance and population structure on genomic and machine-learning results. After quality control and relatedness filtering, 48,902 SNPs from 722 sheep representing 14 populations were retained. The focal dataset remained unbalanced (KIR, n = 187; ALA, n = 10; GIS, n = 15; LOC, n = 43), so balanced resampling was used for sensitivity analyses. KIR showed lower observed heterozygosity (Ho = 0.249) than expected heterozygosity (He = 0.355), with an elevated F IS (0.296), while nucleotide diversity remained relatively high (π = 0.356). Balanced resampling reproduced this contrast, whereas the overall ROH burden in KIR was modest (mean F ROH = 0.009). PCA supported genetic differentiation between KIR and the three Kyrgyzstan local populations, while complementary ancestry and gene-flow analyses indicated regional genetic connectivity. Balanced pairwise F ST estimates between KIR and ALA, GIS, and LOC were 0.099, 0.092, and 0.098, respectively. Intersecting low KIR nucleotide diversity with elevated pairwise FST identified 395 candidate genes, including CAMK 2A, CACNA 1C, EPHB 1, ROBO 1, GRID 1, GRID 2, and RXFP 2; however, no GO or KEGG term remained significant after multiple-testing correction. Machine-learning classifiers showed near-perfect discrimination in the full dataset, but performance and feature stability decreased under balanced resampling, and no prioritized SNP remained significant after PC-adjusted association testing at FDR < 0.01. Chinese Kirghiz sheep are genomically differentiated from neighboring Kyrgyzstan local sheep despite evidence of regional genetic connectivity. Sensitivity analyses indicate that sample imbalance and population structure can strongly influence apparent machine-learning performance and marker prioritization. The identified loci should therefore be regarded as population-informative candidates for further validation rather than as confirmed adaptive variants.