Predicting protein-protein interaction with interpretable bilinear attention network

Protein-protein interactions (PPIs) play the key roles in myriad biological processes, helping to understand the protein function and disease pathology. Identification of PPIs and their interaction types through wet experimental methods are costly and time-consuming. Therefore, some computational methods (e.g., sequence-based deep learning method) have been proposed to predict PPIs. However, these methods predominantly focus on protein sequence information, neglecting the protein structure information, while the protein structure is closely related to its function. In addition, current PPI prediction methods that introduce the protein structure information use independent encoders to learn the sequence and structure representations from protein sequences and structures, respectively, without explicitly learn the important local interaction representation of two proteins, making the prediction results hard to interpret.

Considering that current protein structure prediction methods (e.g., AlphaFold2) can accurately predict protein 3D structures and also provide a large number of protein 3D structures, here we present a novel end-to-end framework (called PPI-BAN) to predict PPIs and their interaction types by integrating protein sequence information and 3D structure information. PPI-BAN uses one-dimensional convolution operation (Conv1D) to extract the protein sequence features, employes GeomEtry-Aware Relational Graph Neural Network (GearNet) to learn protein 3D structure features, and adopts a deep bilinear attention network (BAN) to learn the joint features between one protein sequence and its 3D structure. The sequence features, structure features and joint features are concatenated to fed into a fully connected network for predicting PPIs and their interaction types.

Experimental results show that PPI-BAN achieves the best overall performance against other state-of-the-art methods.

PPI-BAN can effectively predict PPIs and their interaction types, and identify the significant interaction sites by computing attention weight maps and mapping them to specific amino acid residues.