High-resolution population structure inference using genome-wide short tandem repeat variations

Abstract

Short tandem repeats (STRs) are a major source of genetic variation, yet their potential for genome-wide population structure inference remains underexplored. Here we present a multi-modal framework for STR-based population inference, integrating unsupervised clustering, supervised population assignment, and a novel admixture inference model, Directional Non-negative Matrix Factorization (dNMF). Applying this framework to thousands of genomes from multiple global cohorts, we first demonstrate that genome-wide STR variations provide substantially finer resolution of human population structure than single-nucleotide polymorphisms (SNPs), particularly at regional levels. The dNMF model estimates ancestry coefficients under the hypothesis that true ancestral populations are consistently encoded in the bidirectional mutation dynamics of STRs. Population structures inferred by dNMF are fine-grained, reproducible across datasets, and robust to technical artifacts. Motif-specific analyses further reveal directionbiased mutational tendencies and show that distinct STR motif classes encode complementary layers of population structure at different evolutionary scales. These results establish STRs as powerful and biologically interpretable markers for population structure inference, offering a mutation-aware perspective that complements traditional SNP-based frameworks and refines understanding of human demographic history.