Transfer-DDG: Prediction of protein-protein binding affinity changes with mutations based on large pre-trained model transfer learning

The majority of proteins in organisms function through protein-protein interactions (PPIs). Moreover, specific amino acid mutations on these proteins may alter their functionality, which in turn may impact their interactions with other proteins. The investigation of the effect of mutations on protein interactions can be possible by using the numerical index of change in protein-protein binding affinity (ΔΔG). The conventional lab experiments to determine ΔΔG values are both inefficient and costly. In recent years, machine learning and deep learning techniques have been developed for rapid prediction of ΔΔG values. However, many of these approaches fail to extract the deep information embedded in protein amino acid sequences, resulting in unstable prediction results, and they do not generalize well to different types of protein ΔΔG prediction tasks. In this paper, we introduce a deep learning framework for predicting ΔΔG based on protein amino acid sequences, named TransferDDG. The framework utilizes large pre-trained models to learn features of individual amino acid levels, as well as a BiLSTM module to learn features of amino acid sequence levels. This enables the extraction of deep semantic information in protein sequences across multiple dimensions and channels. The model has achieved good results on both SKP1102s and SKP1400m datasets containing single amino acid mutations, surpassing other baseline models.