Improving the Accuracy of Distance-Based Protein–Ligand Binding Affinity Prediction Using Linear Regression and Artificial Neural Network ¹

In the traditional scoring functions for protein–ligand binding affinity prediction, the energies of the electrostatic and van der Waals interactions were evaluated (or restricted) by the mathematical expressions of and , respectively. In comparison, the power exponents of distance-based variables as adopted in the present study are not restricted as those in traditional energy terms for atomic interactions. The distance-based variables were integrated using linear regression and artificial neural network to predict the protein–ligand binding affinity or binding energy. The training of the linear, neural network and mixed models was based on the newest data in PDBbind, i.e ., PDBbind (v.2024). Estimated according to Pearson’s correlation coefficient ( R ), the best performances of the linear models are 0.700 < R ≤ 0.800 with the high-quality affinity data, and those of the neural network-based mixed models are 0.800 ≤ R < 0.900 with the same data. The predictive powers of the best models developed in this study are superior to the sophisticated linear and machine learning-based scoring functions developed before. The results suggest that the distance-based variables with appropriate power exponents may have the ability to improve the prediction of protein–ligand binding affinity with high accuracy.