Moderate-density SNPs combined with machine learning method driven kinship identification and East Asian subpopulations discrimination analysis.

Moderate-density single nucleotide polymorphisms (SNPs) (nearly 2000 SNPs) have shown great forensic application value in Chinese Han. However, the forensic efficacy of these SNPs still requires validation in more populations. In this study, we evaluated the performance of these moderate-density SNPs combined with machine learning (ML) method for identifying first- to fourth-degree kinships in Chinese Yunnan Zhuang (CYZ) group. Moreover, we integrated genetic data from 135 populations spanning nine geographic regions obtained from public datasets to conduct comprehensive population genetic analyses. The present results showed that these SNPs were effective in identifying first- to third-degree kinships from unrelated individuals. The density curves of Log10 likelihood ratio (Log10LR) and cumulative identity by state (CIBS) for first- to second-degree kinships and unrelated individuals were completely separated. When the Log10LR thresholds were set at -4 and 4 to distinguish the third-degree kinship from unrelated individuals, the system power was 86.40% with error rate of 0. The K-Nearest Neighbor model using Log10LR and CIBS feature sets performed best in classifying first- to fourth-degree kinships, with F1 score of 0.9781. Population genetic analyses revealed that CYZ group exhibited the closest genetic affinity with geographically adjacent populations, particularly the southern Chinese minorities (Dai, Miao, and Tujia groups) and the Kinh from Vietnam in Southeast Asia. Principal component, phylogenetic tree and ADMIXTURE analyses revealed significant differences in the genetic structures among East Asian subpopulations, especially the minorities from southern and northern China, suggesting that these SNPs had potential to distinguish these subpopulations. Multinomial LASSO analysis was used to screen 688 SNPs to discriminate five East Asian subpopulations. Using these selected SNPs, a partial least squares-discriminant analysis model based on hierarchical classification approach and an Elastic Net model based on flat multi-classification approach both achieved overall accuracies above 0.9 in classifying five East Asian subpopulations. This study demonstrated that the combination of moderate-density SNPs with ML method could provide an effective solution for kinship identification and fine-scale discrimination of East Asian subpopulations.