Learning strategies can be used to improve the efficiency of virtual screening of very large databases. In these strategies new compounds to be screened are selected on the basis of the results obtained in previous stages, even if truly good ligands have not yet been identified. This approach requires that the scoring function used correctly predicts the energy and geometry of suboptimal complexes, i.e. weak complexes that are not the final solution of the screening but help direct the search toward the most productive regions of chemical space. We show that a small modification in the treatment of the solvation of polar atoms corrects the tendency of the original Autodock 3.0 scoring function to bury ligand polar atoms away from solvent, even if no complementary groups are present in the target and improves the performance of Autodock 3.0 and 4.0 in reproducing the experimental docking energies of weak complexes, resembling the suboptimal complexes encountered in the intermediate stages of virtual screening.