RARE: evolutionary feature engineering for rare-variant bin discovery

Features with rare states, such as rare genetic variants, pose a significant challenge for both statistical and machine learning analyses due to limited detection power and uncertainty surrounding the nature of their role (e.g., additive, heterogeneous, or epistatic) in predicting outcomes such as disease phenotype. Rare variant 'bins' (RVBs) hold the potential to increase association detection power. However, previously proposed binning approaches relied on prior-knowledge assumptions, instead of data-driven techniques, and ignored the potential for multivariate interactions. We present the Relevant Association Rare-variant-bin Evolver (RARE), the first evolutionary algorithm for automatically constructing and evaluating RVBs with either univariate or epistatic associations. We evaluate RARE's ability to correctly bin simulated rare-variant associations over a variety of algorithmic and dataset scenarios. Specifically, we examine (1) ability to detect RVBs of univariate effect (with or without noise), (2) using fixed vs. adaptable bins sizes, (3) employing expert knowledge to initialize bins, and (4) ability to detect RVBs interacting with a separate common variant. We present preliminary results demonstrating the feasibility, efficacy, and limitations of this proposed rare-variant feature engineering algorithm.