Hierarchically Gated Multiplex Heterogeneous Graph Neural Network with Self-Distillation for Protein Stability Prediction.

Predicting the impact of protein residue mutations on thermal stability is critical for protein engineering and for understanding how mutations contribute to cancer development. Recent advances in deep learning demonstrate great promise in the prediction of mutation-induced protein stability changes; however, current approaches face several drawbacks to be solved: 1) neglect the inherent heterogeneity of relations among residues, 2) lack of information-filtering from noisy context, and 3) insufficient supervision of sparse and weak signals of mutation. To address these issues, we propose a protein stability prediction model named HierGate, which is featured by a hierarchically gated multiplex heterogeneous graph neural network with self-distillation mechanism. It is designed to capture the multiplex heterogeneous relations among residues; on the other hand, hierarchical gating mechanism is introduced to suppress contextual noise upon heterogeneous neighbor aggregation. In addition, self-distillation strategy is leveraged to amplify the weak mutation signal. Comprehensive evaluations on eight benchmark datasets demonstrate that the proposed method outperforms advanced protein stability prediction approaches, with an improvement of up to 13% over the second-best baseline on certain datasets. We also demonstrate the practical capability of HierGate in investigating the underlying mechanisms of cancer development and in guiding directed evolution in protein engineering.