ProAffinity-GNN:\nA Novel Approach to Structure-Based\nProtein–Protein Binding Affinity Prediction via a Curated Data\nSet and Graph Neural Networks

Protein–protein\ninteractions (PPIs) are crucial for understanding\nbiological processes and disease mechanisms, contributing significantly\nto advances in protein engineering and drug discovery. The accurate\ndetermination of binding affinities, essential for decoding PPIs,\nfaces challenges due to the substantial time and financial costs involved\nin experimental and theoretical methods. This situation underscores\nthe urgent need for more effective and precise methodologies for predicting\nbinding affinity. Despite the abundance of research on PPI modeling,\nthe field of quantitative binding affinity prediction remains underexplored,\nmainly due to a lack of comprehensive data. This study seeks to address\nthese needs by manually curating pairwise interaction labels on available\n3D structures of protein complexes, with experimentally determined\nbinding affinities, creating the largest data set for structure-based\npairwise protein interaction with binding affinity to date. Subsequently,\nwe introduce ProAffinity-GNN, a novel deep learning framework using\nprotein language model and graph neural network (GNN) to improve the\naccuracy of prediction of structure-based protein–protein binding\naffinities. The evaluation results across several benchmark test sets\nand an additional case study demonstrate that ProAffinity-GNN not\nonly outperforms existing models in terms of accuracy but also shows\nstrong generalization capabilities.