Gene prioritization in GWAS loci using multimodal evidence

Genome-wide association studies (GWAS) yield large numbers of genetic loci associated with traits and diseases. Predicting the effector genes that mediate these locus associations remains challenging. Here we present the FLAMES framework, which predicts the most likely effector gene in a locus. FLAMES integrates machine learning predictions from biological data linking single nucleotide polymorphisms (SNPs) to genes with GWAS-wide convergence of gene interactions. We benchmark FLAMES on gene-locus pairs derived by expert curation, rare variant implication, and domain knowledge of molecular traits. We demonstrate that combining SNP-based and convergence-based modalities outperforms prioritization strategies using a single line of evidence. Applying FLAMES, we resolve the FSHB locus in the GWAS for dizygotic twinning and further leverage this framework to find novel schizophrenia risk genes that converge with rare coding evidence and are relevant in different stages of life.