Improving protein-protein interaction site prediction using graph neural network and structure profiles

Abstract

Protein-protein interactions (PPIs) play a pivotal role in numerous biological processes. Accurate identification of the amino acid residues involved in these interactions is essential for understanding the functional mechanisms of proteins. To effectively integrate both structure and sequence information, we propose a new interaction site predictor, TargetPPI, which leverages bidirectional long short-term memory networks (Bi-LSTM), convolutional neural networks (CNN), and Edge Aggregation through Graph Attention layers with Node Similarity (EGR-NS) neural networks. In TargetPPI, CNN and Bi-LSTM are first employed to extract the global and local feature information, respectively. The combination of global and local features is then used as node embeddings in the graph derived from the protein structure. We have also extracted six discriminative structural features as edge features in the graph. Additionally, a mean ensemble strategy is used to integrate multiple prediction models with diverse model parameters into the final model, resulting in more accurate PPIs prediction performance. Benchmarked results on seven independent testing datasets demonstrate that, compared to most of the state-of-the-art methods, TargetPPI achieves higher accuracy, precision, and Matthews Correlation Coefficient (MCC) values on average, specifically, 84.3 %, 57.6 %, and 0.383, respectively. The source code of TargetPPI is freely available at https://github.com/bukkeshuo/TargetPPI.