Isoform-level transcriptome-wide association uncovers extensive novel genetic risk mechanisms for neuropsychiatric disorders in the human brain

Abstract

Integrative methods, like colocalization and transcriptome-wide association studies (TWAS), identify transcriptomic mechanisms at only a fraction of trait-associated genetic loci from genome-wide association studies (GWAS). Here, we show that a reliance on reference functional genomics panels of only total gene expression greatly contributes to this reduced discovery. This is particularly relevant for neuropsychiatric traits, as the brain expresses extensive, complex, and unique alternative splicing patterns giving rise to multiple genetically-regulated transcript-isoforms per gene. Integrating highly correlated transcript-isoform expression with GWAS requires methodological innovations. We introduce isoTWAS, a multivariate framework to integrate genetics, isoform-level expression, and phenotypic associations in a step-wise testing framework, and evaluate it using data from the Genotype-Tissue Expression (GTEx) Project, PsychENCODE Consortium, and other sources. isoTWAS shows three main advantages. First, joint, multivariate modeling of isoform expression from cis-window SNPs improves prediction by [~]1.8-2.4 fold, compared to univariate modeling. Second, compared to gene-level TWAS, these improvements in prediction lead to [~]1.9-2.5-fold increase in the number of testable genes and a median of 25-70% increase in cross-validated prediction of total gene expression, with the added ability to jointly capture expression and splicing mechanisms. In external validation, isoform-centric models predicted gene expression at percent variance explained >1% for 50% more genes than gene-centric models. Third, across 15 neuropsychiatric traits, isoTWAS increased discovery of trait associations within GWAS loci over TWAS, capturing [~]60% more unique loci and 95% of loci detected by TWAS. Results from extensive simulations showed no increase in false discovery rate and reinforce isoTWASs advantages in prediction and trait mapping power over TWAS, especially when genetic effects on expression vary across isoforms of the same gene. We illustrate multiple biologically-relevant isoTWAS-identified trait associations undetectable by gene-level methods, including isoforms of AKT3, CUL3, and HSPD1 with schizophrenia risk, and PCLO with multiple disorders. The isoTWAS framework addresses an unmet need to consider the transcriptome on the transcript-isoform level to increase discovery of trait associations, especially for brain-relevant traits.