Transcriptome-wide and stratified genomic structural equation modeling identify neurobiological pathways shared across diverse cognitive traits

Functional genomic methods are needed that consider multiple genetically correlated traits. Here we develop and validate Transcriptome-wide Structural Equation Modeling (T-SEM), a multivariate method for studying the effects of tissue-specific gene expression across genetically overlapping traits. T-SEM allows for modeling effects on broad dimensions spanning constellations of traits, while safeguarding against false positives that can arise when effects of gene expression are specific to a subset of traits. We apply T-SEM to investigate the biological mechanisms shared across seven distinct cognitive traits (N = 11,263–331,679), as indexed by a general dimension of genetic sharing (g). We identify 184 genes whose tissue-specific expression is associated with g, including 10 genes not identified in univariate analysis for the individual cognitive traits for any tissue type, and three genes whose expression explained a significant portion of the genetic sharing across g and different subclusters of psychiatric disorders. We go on to apply Stratified Genomic SEM to identify enrichment for g within 28 functional categories. This includes categories indexing the intersection of protein-truncating variant intolerant (PI) genes and specific neuronal cell types, which we also find to be enriched for the genetic covariance between g and a psychotic disorders factor.