Genome-wide association studies have transformed our understanding of complex trait genetics, but most statistical methods assume relatively homogeneous ancestry. This leaves admixed individuals, whose genomes are mosaics of ancestry tracts from multiple continental populations, frequently excluded from analysis due to concerns over population stratification. My lab addresses this gap by developing and applying tools leveraging local ancestry: a technique that assigns ancestral origin to each segment of the genome, rather than summarizing ancestry for individuals using a single label.

I will present work spanning three areas. First, a simulation-based framework establishing best practices for local ancestry inference across reference panel compositions, admixture demographics, and genotype discovery approaches. Second, our recently published method Tractor-Mix, a local ancestry-informed mixed model that extends our earlier Tractor framework to enable well-calibrated GWAS in admixed cohorts with relatedness. We demonstrate Tractor-Mix’s performance across multiple large datasets including the UK Biobank, Yale-Penn cohort, and Mexico City Prospective Study, finding novel loci missed by traditional methods and better determining the ancestry driving unique signals. Third, I will describe our application of local ancestry inference across more than 140,000 genomes in gnomAD, which substantially refines ancestry-specific allele frequency estimates with direct implications for clinical variant interpretation. Together, this work illustrates how ancestry-aware statistical methods can improve genetic discovery, risk prediction, and variant interpretation across the full range of human genetic diversity.