Multivariate variance components analysis uncovers genetic architecture of brain isoform expression and novel psychiatric disease mechanisms

Multivariate variance components linear mixed models are fundamental statistical models in quantitative genetics, widely used to quantify SNP-based heritability (h2SNP) and genetic correlation (rg) across complex traits. However, maximum likelihood estimation of multivariate variance components models remains numerically challenging when the number of traits and variance components are both greater than two. To address this critical gap, here we introduce a novel statistical method for fitting multivariate variance components models. This method improves on existing methods by allowing for arbitrary number of traits and/or variance components. We illustrate the utility of our method by characterizing for the first time the genetic architecture of isoform expression in the human brain, modeling up to 23 isoforms jointly across [~]900 individuals within PsychENCODE. We find a significant proportion of isoforms to be under genetic control (17,721 of 93,293 isoforms) with substantial shared genetic effects among local (or cis-) relative to distal (or trans-) genetic variants (median rg,cis and rg,trans = 0.31 and 0.06). Importantly, we find that 11.6% of brain-expressed genes (2,900 genes) are heritable only at the isoform-level. Integrating these isoform-specific genetic signals with psychiatric GWAS signals uncovers previously hidden psychiatric disease mechanisms. Specifically, we highlight reduced expression of a specific XRN2 isoform as the underlying driver of the strongest GWAS signal for autism spectrum disorder. Overall, our method for fitting multivariate variance components models is flexible, widely applicable, and is implemented in the Julia programming language and available online.