To test this possibility, we imputed gene expression from genotypes using an elastic net model and examined the correlation between the observed genetic ancestry effect from our ancestry differential expression analysis and the predicted genetic ancestry effect computed from the predicted expression. eGenes showed higher prediction accuracy than non-eGenes, with eGenes exhibiting an ancestry difference in gene expression showing a stronger genetic component (higher R2) across brain regions (Supplementary Fig. 25). Furthermore, the imputed gene expression explained an average of 59.5%, 58.7%, 56.8% and 56.8% of the variance in genetic ancestry effect sizes across the caudate nucleus, dentate gyrus, DLPFC and hippocampus, respectively (Fig. 4e). This variance was generally increased at the isoform level (transcript R2 = 50.8% ± 7.0%; exon R2 = 61.6% ± 4.1%; and junction R2 = 62.6% ± 5.1%; Supplementary Fig. 26). In contrast, the genetic variant for the top main effect eQTL associated with these genes explained on average approximately 20% of the variance in genetic ancestry effect sizes with a proportion similar to the isoform level (Supplementary Fig. 27). Thus, genetic variants contributed to nearly 60% of the observed genetic ancestry in gene expression; variant effects on alternative splicing were even greater.
