Purpose

Understanding the behaviour of deep learning (DL) models when applied to genotype–phenotype mapping has been a recent focus at Arcadia [1][2]. To date, our efforts have largely focused on exploring the predictive performance of various types of models (but see [3]). In this notebook, however, we extend our work to trying to understand how we might extract functional information from trained DL models. We apply a recently described method for obtaining equivalent linear representations of DL models to assess feature importance per input sample. We show that under certain conditions, the distribution of feature importance values across samples can be used to infer familiar quantitative genetic parameters such as additive effect sizes, epistatic interactions, and genetic variance components, allowing for the use of DL models in tasks such as genomic target identification and interaction mapping.