Abstract:
Conventional methods for predicting protein–protein interactions (PPIs) often depend on intricate amino acid-level data obtained from both sequences and structures. Although effective, such methods typically require high-definition information and considerable computational power. Here, we present CurvePotGCN, an innovative graph convolutional neural network that predicts PPIs through a simplified physicochemical model of protein surfaces. Our approach represents proteins as graphs wherein nodes symbolize surface clusters defined by geometric curvature and electrostatic potential, focusing exclusively on these fundamental physicochemical features rather than evolutionary conservation or complex machine learning representations. This model is built on the principle that complementary shape and electrostatic potential at the protein–protein interface are primary determinants of whether two proteins interact. CurvePotGCN achieved a predictive performance of 98% area under the receiver operating characteristic curve for human PPI and 89% for yeast PPI. Upon benchmarking, CurvePotGCN showed superior performance against contemporary methods, highlighting the effectiveness of using reduced, physicochemically based models for PPI prediction. Our study demonstrates that using biophysical properties as features can provide competitive performance to more complex representation schemes, enhancing computational efficiency while maintaining predictive accuracy.