Deriving Mendelian Randomization–Based Causal Networks of Brain Imaging Phenotypes and Bipolar Disorder

Background: Neuroanatomical variation in individuals with bipolar disorder (BD) has been described previously in observational studies. However, the causal dynamics of these relationships remain unexplored. Methods: We performed Mendelian randomization (MR) of 297 structural and functional neuroimaging phenotypes from the UK Biobank and BD using genome-wide association study summary statistics. We carried out a suite of sensitivity analyses and identified phenotypic categories with the greatest effect on BD. We applied a novel inverse sparse regression model that accounts for covariance between sets of correlated effects to estimate direct causal effects (DCEs), which represent the effect of one phenotype conditional on all other effects. We used DCE weights to create causal scores for BD using neuroimaging data from 3 clinical cohorts. Results: We found 28 significant causal relationship pairs after multiple testing correction containing BD as a term, 27 of which described neuroimaging phenotype effects on BD. White matter tract phenotypes had larger absolute effects on BD than vice versa in MR tests and estimated DCE solutions. We found that white matter phenotypes had significantly larger out-degrees than non–white matter tract phenotypes across network solutions. A causal score constructed using neuroimaging causal estimates was a significant predictor of BD in an adolescent cohort (odds ratio = 0.79). Conclusions: MR analyses suggest that neuroanatomical variation, specifically in white matter tracts such as the longitudinal fasciculi, is likely a cause rather than a consequence of BD. Verification of estimated causal relationships requires replication and triangulation of evidence approaches using other study designs.