Statistical and functional fine-mapping as a powerful tool to unravel the genetic etiology of bipolar disorder

The Psychiatric Genomics Consortium has published the largest GWAS of bipolar disorder (BD) (41,917 cases and 371,549 controls) which identified 64 BD risk loci. However, each locus encompasses many associated genetic variants and genes, and those causal for BD remain largely unknown. Fine-mapping is a rapid, scalable and cost-effective approach to computationally prioritize likely causal genes and SNPs for functional laboratory experiments. We fine-mapped the 64 BD risk loci identified by the largest BD GWAS to date, as published by the Psychiatric Genomics Consortium (PGC3). Stepwise conditional analysis of each locus using GCTA-COJO and the Haplotype Reference Consortium (HRC) LD reference panel, indicated one association signal per locus. To allow exploratory analysis, we selected two different numbers of maximum causal variants per each locus (–max-num-causal =1 and –max-num-causal = 5). Subsequently, each locus was fine-mapped using a suite of Bayesian fine-mapping methods: FINEMAP, SuSiE, PolyFun-FINEMAP and Polyfun-SuSiE. Each fine-mapping method generates SNP-wise posterior probabilities of causality (PP), and a 95% credible set (CS), containing the candidate causal SNPs. Consensus SNPs were defined as those in the 95% CS from at least two methods. Published functional annotations from the UK Biobank (UKB) baseline-LF model were incorporated into PolyFun to compute prior causal probabilities (priors). To thoroughly examine the impact of LD on fine-mapping, results were compared using LD estimates from the HRC, UKB, and PGC3 BD cohorts. We identified 28 genetic variants with a >50% probability of being causal for BD (PIP > 0.50), based on convergence of evidence across fine-mapping methods ('consensus SNPs'). The size of CSs was assessed across the different fine-mapping methods, LD panels and fine-mapping ‘windows’. Most consensus SNPs were included in 95% CSs of a small SNP size. Of the consensus SNPs, 3 were missense variants within the SCN2A, FKBP2 and TRPT1 genes, and 13 overlapped with promoters or enhancers of gene expression in specific brain cell-types. Using expression and splicing quantitative trait loci (eQTLs and sQTLs) from PsychENCODE, CommonMind and ROSMAP bulk brain tissue (dorsolateral prefrontal cortex), we detected eQTLs and sQTLs which colocalized with BD loci at the FURIN, FKBP2, SYNE1, THRA, SRPK2 and RPL13 genes. There was convergent evidence from fine-mapping and studies of rare genetic variation in BD, at the THSD7A, SCN2A and CACNA1B genes. Finally, we demonstrated that integrating information from fine-mapping improved the R2 variance explained (liability scale) of polygenic risk scores by up to 1.8%. This comprehensive pipeline, implemented through Snakemake, will be made publicly available for the rapid, reproducible, and scalable fine-mapping of GWAS risk loci across psychiatric disorders, hence prioritizing genetic variants and genes for further experimental validation.