Scalable high resolution ancestry deconvolution for genomic data

Abstract

As genome-wide association studies and genetic risk prediction models extend to globally diverse and admixed biobanks, accurate, scalable ancestry deconvolution, also called local ancestry inference (LAI), has become crucial. LAI assigns ancestry to each genomic segment within an individual, enabling studies of population history and ancestry-associated haplotypic effects. Existing LAI methods scale poorly to biobank-scale data, to the distant past, and to large numbers of ancestries. Here, we introduce several independent LAI methods implemented in the Gnomix software suite, achieving higher accuracy and faster computational performance than all existing approaches and with portable models that can be shared without exposing individual-level training data. Gnomix is paired with Gnofix, a swift, scalable phase correction counterpart. We demonstrate performance on worldwide whole-genome data from humans and canids, leveraging high-resolution accuracy to localise ancient New World haplotypes in the Xoloitzcuintli, dating back over 100 generations. Code is available at https://github.com/AI-sandbox/gnomix. The authors present Gnomix, a local ancestry framework that delivers leading accuracy across diverse admixed datasets on both whole-genome and array data with high efficiency, together with Gnofix, its fast phasing-error correction counterpart.