Advancing protein-ligand affinity prediction: an enhanced Convex-PL framework with three-pronged optimization strategies

Protein-ligand interactions are pivotal in drug discovery and the study of biological macromolecules. Accurately predicting protein-ligand binding affinity accelerates drug development and reduces experimental costs. Existing scoring functions, such as Convex-PL, have made progress but still face challenges in robustness, feature selection, and ranking quality. This paper presents an improved Convex-PL model that addresses these challenges through three key enhancements: Adaptive Hinge Loss for smoother optimization landscapes, Elastic Net Regularization for better feature selection, and Listwise Ranking Loss for improved efficiency in ranking tasks. The proposed model was evaluated on the PDBbind v2016 dataset. Results indicate significant improvements in prediction accuracy and model efficiency. The adaptive hinge loss primarily enhances the smoothness of the optimization function, the elastic net regularization effectively compresses the feature space, and the listwise ranking loss optimizes the ranking quality. Overall, the optimized Convex-PL model demonstrates superior performance, providing a robust tool for protein-ligand interaction studies.