Disentangling the contribution of each descriptive characteristic of every single mutation to its functional effects

Abstract

Artificial intelligence (AI) based methods for mutational effects predictions are improving in accuracy, because of the exhaustive experimental data they are trained on, and advances in algorithms. As the prediction quality improves, the next natural question to ask countering the 'black box' image of AI is if the predictions can be interpreted. We applied one of the approaches developed in the field of explainable AI, to decipher the factors contributing to the changes in cellular fitness and protein solubility arising from mutations. Juxtaposing the two observations and focusing on the individual factors uncovers the contributions and quantifies the intuitions about how different factors such as conservation, distance from the catalytic site affect fitness and solubility. Embedding interpretability along with the prediction algorithms will enable transparency and inspire confidence into models as well as contribute to the understanding of how mutations affect proteins.