We present the third generation of our scoring function for the prediction of protein–ligand binding free energy. This function is now a hybrid between a knowledge-based potential and an empirical function. We constructed a diversified set of ∼1000 complexes from the PDBBinding-CN database for the training of the function, and we show that this number of complexes generates enough data to build the potential. The occurrence of 420 different types of atomic pairwise interactions is computed in up to five different ranges of distances to derive the knowledge-based part. All of the parameters were optimized, and we were able to considerably improve the accuracy of the scoring function with a Pearson correlation coefficient against experimental binding free energies of up to 0.57, which ranks our new scoring function as one of the best currently available and the second-best in terms of standard deviation (SD = 1.68 kcal/mol). The function was then further improved by inclusion of different terms taking into account repulsion and loss of entropy upon binding, and we show that it is capable of recovering native binding poses up to 80% of the time. All of the programs, tools, and protein sets are released in the Supporting Information or as open-source programs.