AI tool adjusts for ancestral bias in genetic data

Human ancestry has a considerable impact on gene expression, but genomic datasets for disease analysis severely underrepresent non-European populations, thereby limiting the advancement of precision medicine. In a paper in Nature Communications, Smith et al. introduce a machine learning tool to mitigate the effects of ancestral bias in transcriptomic data.

The tool, called PhyloFrame, creates ancestry-aware signatures of disease by integrating population genomics data with smaller, disease-relevant training datasets. PhyloFrame uses a logistic regression model with LASSO penalty to obtain an initial set of disease-relevant genes. It then uses population genomics data to help compensate for data distribution shifts caused by human ancestry differences. In short, PhyloFrame projects the initial disease signature onto a functional interaction network, extending the network to include the first and second neighbors of each signature gene. This new set is then filtered by a statistic defined as enhanced allele frequency (EAF) — which captures population-specific allelic enrichment in healthy tissue — to identify ancestrally diverse genes that interact with the original signature. From each ancestry, a selected subset of genes with high EAF and gene expression variability in the training data are added to the PhyloFrame signature. Retraining the model with the forced inclusion of these equitable genes results in a signature of disease that generalizes to all populations, even if not represented in the training data.