Improving Inverse Folding models at Protein Stability Prediction without additional Training or Data

Abstract

AbstractDeep learning protein sequence models have shown outstanding performance at de novo protein design and variant effect prediction. We substantially improve performance without further training or use of additional experimental data by introducing a second term derived from the models themselves which align outputs for the task of stability prediction. On a task to predict variants which increase protein stability the absolute success probabilities of ProteinMPNN and ESMifare improved by 11% and 5% respectively. We term these models ProteinMPNN-ddG and ESMif-ddG.