Hypothesis-driven interpretable neural network for interactions between genes

Mechanistic models of genetic interactions are rarely feasible due to a lack of information and computational challenges. Alternatively, machine learning (ML) approaches may predict gene interactions if provided with enough data but they lack interpretability. Here, we propose an ML approach for interpretable genotype-fitness mapping, the Direct-Latent Interpretable Model (D-LIM). The neural network is built on a strong hypothesis: mutations in different genes cause independent effects in phenotypes, which then interact via non-linear relationships to determine fitness. D-LIM predicts genotype-fitness maps for combinations of mutations in multiple genes with state-of-the-art accuracy, showing the validity of the hypothesis in the case of a deep mutational scanning in a metabolic pathway. The hypothesisdriven structure of D-LIM offers interpretable features reminiscent of mechanistic models: the inference of phenotypes, fitness extrapolation outside of the data domain, and enhanced prediction in low-data regimes by the integration of prior knowledge.