Valid inference for machine learning-assisted GWAS

Abstract

Machine learning (ML) has revolutionized analytical strategies in almost all scientific disciplines including human genetics and genomics. Due to challenges in sample collection and precise phenotyping, ML-assisted genome-wide association study (GWAS) which uses sophisticated ML to impute phenotypes and then performs GWAS on imputed outcomes has quickly gained popularity in complex trait genetics research. However, the validity of associations identified from ML-assisted GWAS has not been carefully evaluated. In this study, we report pervasive risks for false positive associations in ML-assisted GWAS, and introduce POP-GWAS, a novel statistical framework that reimagines GWAS on ML-imputed outcomes. POP-GWAS provides valid statistical inference irrespective of the quality of imputation or variables and algorithms used for imputation. It also only requires GWAS summary statistics as input. We employed POP-GWAS to perform the largest GWAS of bone mineral density (BMD) derived from dual-energy X-ray absorptiometry imaging at 14 skeletal sites, identifying 89 novel loci reaching genome-wide significance and revealing skeletal site-specific genetic architecture of BMD. Our framework may fundamentally reshape the analytical strategies in future ML-assisted GWAS.

Detecting potential gene-environment interaction effects involved in behavioural traits across childhood: variance quantitative trait loci mapping in the MoBa cohort

“…Due to the large number of analyses, we performed a complete case analysis for each phenotype. However, novel approaches which enable unbiased GWAS on imputed phenotypes are currently in development 44 . We have analysed a large number of traits, which increases the multiple testing burden.…”