Leveraging a machine learning derived surrogate phenotype to improve power for genome-wide association studies of partially missing phenotypes in population biobanks

Abstract

Within population biobanks, genetic discovery for specialized phenotypes is often limited by incomplete ascertainment. Machine learning (ML) is increasingly used to impute missing phenotypes from surrogate information. However, imputing missing phenotypes can invalidate statistical inference when the imputation model is misspecified, and proxy analysis of the ML-phenotype can introduce spurious associations. To overcome these limitations, we introduce SynSurr, an approach that jointly analyzes a partially missing target phenotype with a "synthetic surrogate", its predicted value from an ML-model. SynSurr estimates the same genetic effect as standard genome-wide association studies (GWAS) of the target phenotype, but improves power provided the synthetic surrogate is correlated with the target. Unlike imputation or proxy analysis, SynSurr does not require that the synthetic surrogate is obtained from a correctly specified generative model. We perform extensive simulations and an ablation analysis to compare SynSurr with existing methods. We also apply SynSurr to empower GWAS of dual-energy x-ray absorptiometry traits within the UK Biobank, leveraging a synthetic surrogate composed of bioelectrical impedance and anthropometric traits.