Surflex-Dock employs an empirically derived scoring function to rank putative protein−ligand interactions by flexible docking of small molecules to proteins of known structure. The scoring function employed by Surflex was developed purely on the basis of positive data, comprising noncovalent protein−ligand complexes with known binding affinities. Consequently, scoring function terms for improper interactions received little weight in parameter estimation, and an ad hoc scheme for avoiding protein−ligand interpenetration was adopted. We present a generalized method for incorporating synthetically generated negative training data, which allows for rigorous estimation of all scoring function parameters. Geometric docking accuracy remained excellent under the new parametrization. In addition, a test of screening utility covering a diverse set of 29 proteins and corresponding ligand sets showed improved performance. Maximal enrichment of true ligands over nonligands exceeded 20-fold in over 80% of cases, with enrichment of greater than 100-fold in over 50% of cases.