Joint Bayesian modelling of molecular QTL and GWAS effects improves polygenic prediction for complex traits

Abstract

Integrating molecular quantitative trait locus (molQTL) data into polygenic prediction offers a promising strategy for improving complex trait prediction. We introduce SBayesCO, a Bayesian framework that jointly models genome-wide association study (GWAS) and molQTL effect sizes by treating the complex trait and molecular phenotypes as genetically correlated traits. SBayesCO estimates genome-wide SNP effects on both phenotypes and supports analyses using individual- or summary-level data. In simulations, SBayesCO consistently outperforms SBayesC, a baseline model based solely on GWAS data, especially when GWAS sample sizes are modest. Applying to 11 blood and immune-related traits using large-scale expression QTLs (eQTLs) and protein QTLs (pQTLs) resources, SBayesCO achieves relative improvements in prediction R2 of up to 6.3% with pQTLs and 5.3% with eQTLs compared to SBayesC, with similar gains relative to modelling molQTLs as binary annotations (SBayesCC). SBayesCO also improves SNP prioritization by concentrating posterior inclusion probabilities on regulatory variants. These results demonstrate the value of modelling quantitative molQTL effect sizes and provide guidance on how increasingly available functional genomic annotations, including AI-based regulatory effect predictions, can be effectively integrated to improve polygenic prediction.