Interpretable Prediction of Protein Stability Changes upon Mutation by Using Decision Tree

Abstract

For protein stability changes upon mutation, an accurate predictor with linguistic interpretability is beneficial to protein designs. Traditional analysis based on linear correlation between predicted and experimental data reveals their primitive relationships. Recently, some machine learning techniques such as artificial neural network (ANN)-based methods were applied to find an accurate predictor. However, the ANN-predictor without interpretability is insufficient in knowledge discovery. This paper proposes an interpretable predictor using a rule-based decision tree method (named iPTREE) for accurately predicting protein stability changes upon single point mutations. Besides being a sign predictor, iPTREE can be used both as a model for verifying attributes effect, and as a rules miner in the protein stability change study. iPTREE is depending on features including mutation type (deleted and introduced residues), the relative solvent accessibility value (RSA), the experimental conditions (pH and temperature) and the local spatial environment. To evaluate the performance of iPTREE, a thermodynamic dataset consisting of 1615 mutations generated from ProTherm is used. The computer simulation shows that iPTREE has an accurate prediction for the direction of stability changes as high as 87%, which is significantly better than the ANN-predictor for the same features.