PPIM-Struct: Leveraging Structural Embeddings for Predicting Protein-Protein Interaction Modulator Interactions

Abstract

Protein-protein interactions (PPIs) are the basis of many important biological functions. However, previous studies mainly use chemical information such as charge and hydrophobicity to predict PPI modulators, which ignores 2D and 3D structural information that contributes to protein-protein modulator interactions. We develop PPIM-Struct, a structure-based method to predict the interaction between modulators and protein-protein interactions, using GCN-based protein embeddings and comparing different pre-trained modulator embeddings. We found that applying Morgan fingerprints with PPIM-Struct achieves the best F1-score of 82.2% on PPI-modulator interaction prediction, 9.6% and 8.6% higher than using GCN and Transformer embeddings, respectively. The high performance of Morgan fingerprint embeddings highlights the usefulness of a subgraph-level tokenizer for efficient molecular representations. We think this work can be extended to structural representations of similar chemicals, such as peptides, prodrugs, and lipids.