Harnessing angular geometry in deep learning for protein–ligand binding affinity prediction

Protein–ligand binding affinity prediction is essential in structure-based drug design, where binding scores guide the selection of promising candidate ligands. Existing deep learning models often use 3D grids, voxelized complexes, or molecular graphs. These representations are resource-intensive and may not capture specific directional interactions. This paper introduces angular geometric features as key descriptors of binding interactions. Seven types of dihedral angles between protein and ligand atoms are extracted to encode orientation and geometry. A fully connected ensemble network, called the Angle-Aware Predictor (AAP), integrates these features. On CASF-2016, AAP achieves state-of-the-art results with correlation coefficient (R) of 0.872, root mean squared error (RMSE) of 1.072, mean absolute error (MAE) 0.817, standard deviation (SD) of 1.077, and concordance index (CI) of 0.845. On four additional benchmarks, AAP shows consistent improvements ranging from 0.3% to 36%. The angular features are effective, lightweight, and robust descriptors for binding affinity prediction. These results highlight angular geometry as a valuable direction for future structure-based drug discovery. The program and data of AAP are publicly available at https://github.com/juliacse06/AAP . • Angular geometric features are introduced for protein–ligand binding affinity prediction, capturing critical structural interactions beyond conventional distance-based descriptors. • A deep learning–based Angle-aware Affinity Predictor (AAP) is designed to effectively integrate the angular features. • Benchmarking on CASF-2016 and four additional datasets shows AAP’s superior performance with minimal feature complexity.