Explaining potential epistasis in genomic data using symbolic representations of complex black box models

Epistasis, known as the interaction among genetic variants, has long been hypothesized to play a major role in explaining missing heritability. Though recent studies have found many candidate variants demonstrating epistasis signals in the UKBiobank, it remains a controversial question how to interpret the findings. Nonlinear models have shown potential in capturing these signals, but we require additional explanation methods to understand the projected relationships. Here, we utilize symbolic pursuit, a form of symbolic regression that provides a closed-form, interpretable model which generalizes first order explanations. Furthermore, we extend this study by applying Taylor expansions to the model, balancing interpretability with performance while improving its generalizability. We found the method performed reliably and was consistent with other methods across a variety of simulated data. This work contains strong implications for its use on large genomic datasets and its ability to capture nonlinear interactions without prior knowledge of the genetic architecture.

Authors: Anand, A., Anand, P.