Spatial Hierarchical Protein-Protein Interaction Site Prediction Using Squeeze-and-Excitation Capsule Networks

The discovery of protein-protein interaction (PPI) sites is vital for exploring the principles of PPIs. Developing computational approaches to predict PPI sites can effectively compensate for biological experiments, which are mostly time-consuming and vulnerable-to-noise. In recent years, deep learning has been used to predict PPI sites, by considering the contextual information of target amino acid residues and using a local protein sequence to represent the targets. However, traditional deep-learning techniques, e.g., DNNs and CNNs, disregard important spatial hierarchies contained in the features of protein sequences, leading to their failure to effectively distinguish interaction sites from different residue regions. In this work, we design MSE-CapsPPISP, a new deep-learning model to address PPI site prediction with spatial hierarchies. The key idea of MSE-CapsPPISP is to take into account the hierarchical relationships between the features of protein sequences. We characterize the hierarchical relationships by designing a tailored Capsule Network, which is a novel type of neural network with vector neurons. Moreover, to make the network representation more robust, MSE-CapsPPISP uses multi-scale CNNs to extract multi-scale features of protein sequences and Squeeze-and-Excitation blocks to recalibrate the features. The validation results show that MSE-CapsPPISP outperforms the baseline CNNs-based architecture DeepPPISP and other existing competing schemes in the four key metrics of F1, MCC, AUROC, and AUPR.