Learning spatial structures and network correlations improves unknown protein–protein interaction prediction

The mechanism of action of protein-protein interactions(PPIs) is complex, and prediction models have to learn multiple dimensions to achieve excellent generalization performance. Therefore, based on the concept that the spatial structure of proteins is closely related to protein functions and the topological information of PPI networks reflects the correlation between proteins, we combine the protein structure information and the topological information of PPI networks to enhance the prediction performance. We present a new approach, SE3NET-PPI, for multi-type PPI prediction, retrieves protein structure information from the SE (3)-invariant matrix map generated by Alphafold2 and extracts the topological information of the PPI network using a graph neural network in the Siamese architecture. Results showed that our model outperforms several state-of-the-art methods under various dataset partitioning methods, with significant improvement in predicting invisible datasets. For example, in the case of Sring_al1-BFS, the miroc-F1 value of our model reaches 80.28 ± 0.43, compared to 75. 87± 0.37 for GNN-PPI and 62.30 ± 0.41 for PIPR. The implementation and related datasets are available at https://github.coml YY99117/SE3NET-PPI.