Co-expression-based models improve eQTL predictions and highlightnovel transcriptome-wide genes associated with schizophrenia

Abstract

Non-coding genetic variants statistically associated with complex heritability phenotypes are thought to act primarily through transcriptome regulatory mechanisms. Predictions of gene expression in tissue like the human brain traditionally rely primarily on cis-eQTLs. Here, we introduce INGENE and MODULE, trans-eQTLs models designed to enhance the prediction of gene expression by capturing the collective impact of candidate trans-eQTLs acting within co-expression networks. Exploiting RNA-seq data in six post-mortem brain regions (amygdala, caudate nucleus, dorsal/subgenual anterior cingulate cortex, dorsolateral prefrontal cortex, and hippocampus), we validate our models on two testing datasets, demonstrating increased gene predictability compared to both an original cis-based model and to EpiXcan, the leading benchmark in cis-model performance. Integration of cis- and trans-predictions significantly improves gene-level expression imputation (MLE = 0.05) for 18,744 genes across the six brain regions considered. Applying cis and trans models to PGC wave 3 genotypes identifies 766 SCZ-associated genes across brain regions (pFDR < .01), emphasizing the complementary nature of cis and trans predictions in trait association discovery. Of these genes, 641 represent novel transcriptome-wide associations with schizophrenia, highlighting the role of trans-heritability and genetic interactions underlying risk for this disorder, in addition to further supporting 125 previous candidates.