Deep learning-based instrumental variable regression for nonlinear causal inference in Transcriptome-Wide Association Studies

Abstract

Instrumental Variable (IV) regression is a foundational approach to causal inference, particularly in contexts where randomized experiments are not feasible. An important and novel application of IV regression is to Transcriptome-Wide Association Studies (TWAS) to identify causal genes (as exposures) for a trait (as an outcome) using genetic variants as IVs. Despite its widespread use, TWAS methodologies are constrained by several limitations: they predominantly rely on linear models or univariable approaches, thereby neglecting nonlinear gene-trait relationships and the confounding or pleiotropic effects of genetic variants. This gap impedes the accurate and robust identification of causal genetic mechanisms—a critical need for advancing the understanding of complex diseases. In this dissertation, we introduce several novel approaches to address these limitations. First, we propose DeLIVR, a novel deep learning (DL) framework tailored for TWAS applications. DeLIVR improves upon existing DL-based IV regression approaches by estimating a different target function and incorporating a hypothesis-testing framework. Second, we introduce MV-DeLIVR, which extends DeLIVR to multivariate settings, accounting for violations of model assumptions due to genetic pleiotropy. Finally, we explore the application of DeLIVR to imputed traits to enhance statistical power for diseases with limited cases, using Alzheimer’s disease (AD) as a motivating example. The underrepresentation of late-onset AD cases in biobank datasets, due to the younger age of participants relative to the typical onset age, presents a significant challenge for TWAS. In response to this challenge, we train DeLIVR on a large biobank study with imputed AD cases and test the gene-trait associations using a smaller dataset with observed AD status. We validate the accuracy and robustness of our proposed methods through comprehensive simulations and real-data applications.