Combining Causal Inference and Machine Learning for Model-agnostic Discovery in High-dimensional Biology

Abstract

The widespread availability of high-dimensional data has catalyzed the process of biological pattern discovery. Today, the simultaneous screening of anywhere from thousands to millions of biological characteristics (in, e.g., genomics, metabolomics, proteomics, etc.) is commonplace in many experimental settings, making the simultaneous screening of such a large number of characteristics a central problem in computational biology and allied sciences. The information gleaned from such studies promises substantial progress, not only for the basic sciences but also for medicine and public health. While the tools of modern chemical biology and biophysics allow for great precision in probing biological systems at the molecular-cellular level, population biomedical and public health sciences must operate without access to such a fine level of control. Instead, statistical innovations bridge the gap – being used to dissect mechanistic processes and to mitigate the inferential obstacles imposed by the confounding of key relationships in observational (non-randomized) studies. Unfortunately, most off-the-shelf statistical techniques rely on restrictive assumptions – born of mathematical convenience – that invite opportunities for bias due to model misspecification (when the biological process under study fails to obey the assumed mathematical conveniences). Fortunately, model-agnostic inference, which draws on causal inference and semiparametric efficiency theory, provides an avenue for avoiding restrictive modeling assumptions while obtaining robust statistical inference about scientifically relevant parameters. We outline this framework briefly and introduce a model-agnostic approach to biomarker discovery incorporating causal inference-based parameters, leveraging state-of-the-art machine learning for flexible estimation (to mitigate potential for model misspecification), and utilizing variance moderation (to curb Type-I error) to deliver stable inference in high-dimensional settings, even when sample sizes are limited. When paired with domain expertise, this approach reliably identifies biomarkers linked to disease or exposure patterns, yielding insights for future investigations into therapeutics and policy interventions. The approach is implemented in the open-source biotmle R/Bioconductor package (https://bioconductor.org/biotmle). This talk is based on joint work with Alan Hubbard, Mark van der Laan, and Philippe Boileau, described in the pre-print: https://arxiv.org/abs/1710.05451.