MMProtRepPPI: A Multimodal Protein Representation for Predicting Protein–Protein Interactions Using Sequence and Structure

Recently, a range of effective methods have been developed for predicting protein-protein interactions (PPIs). Among them, the methods based on data derived from protein sequences and structures have shown promising results. However, most existing structure-based methods consider only whole-protein structural information, potentially reducing predictive performance by overlooking local regions. To overcome this limitation, we propose a protein representation method that is capable of modeling local regions of protein structure. The local structural regions are then processed by multiple graph neural networks to identify local interaction regions. At the same time, for our PPI prediction model, we incorporate a lightweight neural network to leverage feature-rich sequence embeddings derived from protein language models to further improve the PPI prediction performance. We compare the performance of our methods with that of robust existing sequence-based and structure-based methods. The results show that our methods outperform these methods across all metrics on the Yeast core dataset and most metrics on other benchmark datasets obtained from various organisms. The source code of our method is available at https://gitlab.com/nhanth/MMProtRepPPI.