Exploring the Upper Boundaries of Distance-Based Protein-Ligand Binding Affinity Prediction: Accuracy Improvements via Linear Regression and Artificial Neural Networks

Abstract

The traditional scoring functions for protein-ligand binding affinity prediction usually adopt fixed mathematical formulas to characterize atomic interaction energies, where electrostatic interactions are described by the a/r function and van der Waals interactions by the b/r^6 -c/r^12 function. These fixed power exponents of distance-based variables severely limit the predictive flexibility and generalization of conventional scoring models. To address this limitation, this study proposes an improved prediction strategy with unrestricted power exponents for distance-dependent atomic interaction variables. To explore the upper boundaries of the predictive powers, the optimized distance-based descriptors were integrated with linear regression and artificial neural network algorithms to construct novel models for protein-ligand binding affinity and binding energy prediction. All linear, artificial neural network, and hybrid models in this work were trained and validated on the latest high-quality dataset from PDBbind (version 2024). Quantitative evaluation based on the Pearson's correlation coefficient (R) demonstrated that the optimized linear models achieved a moderate predictive performance with R values ranging from 0.700 to 0.800 on the high-quality affinity dataset, while the artificial neural network-based hybrid models yielded substantially improved accuracy, with R values between 0.800 and 0.900. Comparative analysis verified that the optimal models established in this study outperform previously-reported state-of-the-art linear and machine learning-based scoring functions. Specially, this work indicates that the scoring functions with explicit mathematical expressions can achieve a high accuracy close to the experimental error for affinity prediction in a dataset composed of several thousand protein-ligand complexes. The overall results indicate that rationally-optimized power exponents for distance-based variables can effectively enhance the predictive accuracy of protein-ligand binding affinity prediction, providing a reliable and efficient tool for computational drug discovery.