Representative estimation of allele frequencies is essential for the calculation of robust rarity statistics in forensic DNA casework. Current methods rely on population frequency data derived from samples with predefined population labels, which may be incomplete, uncertain, or reflect social and cultural constructs rather than underlying genetic structure. A data-driven approach to sample selection could complement traditional label-based methods, particularly in large genomic datasets where population membership is ambiguous or unknown.  

In this study, STR genotype data from four continental-level populations in the high coverage Phase 3, 1000 Genomes Project dataset were analyzed using an unsupervised machine learning pipeline to identify genetically structured subsets directly from the data, without reliance on the predefined population labels. Dimensionality reduction followed by Gaussian Mixture Model (GMM) clustering was applied to the mixed dataset in an attempt to recover subsets suitable for population allele frequency estimation. 

To evaluate the suitability of the identified subsets, the 200 individuals with the highest posterior probabilities of cluster membership were selected for allele frequency estimates. These subsets were then subjected to exact tests of Hardy-Weinberg equilibrium, assessing whether the resulting subsets behaved as internally consistent population samples.  

The results demonstrate that unsupervised analysis of large continental-level mixtures of STR genotype data can identify subsets that satisfy the assumptions required for allele frequency estimation. These findings suggest machine learning may provide a complementary, data-driven strategy for population reference sample identification that does not depend on predefined ancestry classifications. 

Author

• Daniel Myers (Syracuse University & New York State Police, United States of America)