Causal considerations can determine the utility of machine learning assisted GWAS
Machine Learning (ML) is increasingly employed to generate phenotypes for genetic discovery, either by imputing existing phenotypes into larger cohorts or by creating novel phenotypes. While these ML-derived phenotypes can significantly increase sample size, and thereby empower genetic discovery, they can also inflate the false discovery rate (FDR). Recent research has focused on developing estimators that leverage both true and machine-learned phenotypes to properly control the type-I error. Our work complements these efforts by exploring how the true positive rate (TPR) and FDR depend on the causal relationships among the inputs to the ML model, the true phenotypes, and the environment. Using a simulation-based framework, we study architectures in which the machine-learned proxy phenotype is derived from biomarkers (i.e. inputs) either causally upstream or downstream of the target phenotype. We show that no inflation of the false discovery rate occurs when the proxy phenotype is generated from upstream biomarkers, but that false discoveries can occur when the proxy phenotype is generated from downstream biomarkers. Next, we show that power to detect variants truly associated with the target phenotype depends on its heritability and correlation with the proxy phenotype. However, the source of the correlation is key to evaluating a proxy phenotype's utility for genetic discovery. We demonstrate that evaluating machine-learned proxy phenotypes using out-of-sample predictive performance (e.g. phenotypic correlation) provides a poor lens on utility. This is because overall predictive performance does not differentiate between genetic and environmental correlation. In addition to parsing these properties of machine-learned phenotypes via simulations, we further illustrate them using real-world data from the UK Biobank.