Essential Regression - a generalizable framework for inferring causal latent factors from multi-omic human datasets

Abstract

High-dimensional cellular and molecular profiling of human samples highlights the need for analytical approaches that can integrate multi-omic datasets to generate predictive biomarkers that are in turn accompanied with strong causal inferences. Current methodologies are challenged by the high dimensionality of the combined datasets, the differences in distributions across the datasets, and their integration in a plausible causal framework, beyond merely correlative biomarkers. Here we present CausER, a first-in-class two-step interpretable machine learning approach for high-dimensional multi-omic datasets, that addresses these problems by identifying latent factors and their cause-effect relationships with the system-wide outcome/property of interest. The first step consists of Essential Regression (ER), a novel data-distribution-free regression model that integrates multi-omic datasets and identifies latent factors significantly associated with an outcome. The second involves probabilistic graphical modeling of the significant latent factors to infer plausible causal associations between them and mechanisms that affect outcomes, thereby significantly moving beyond predictive associative markers. By analyzing varied human immunological multi-omic datasets, we demonstrate that CausER significantly outperforms a wide range of state-of-the-art approaches. It generates novel cellular and molecular predictions in a range of contexts, including immunosenescence and sustained immune dysregulation associated with pre-term birth, that are corroborated by biological findings in model systems.