A comparative binding energy (COMBINE) analysis (Ortiz et al. J. Med. Chem. 1995, 38, 2681-2691) has been performed on a training set of 33 HIV-1 protease inhibitors, and the resulting regression models have been validated using an additional external set of 16 inhibitors. This data set was originally reported by Holloway et al. (J. Med. Chem. 1995, 38, 305-317), who showed the usefulness of molecular mechanics interaction energies for predicting the activity of novel HIV-1 protease inhibitors within the framework of the MM2X force field and linear regression techniques. We first used the AMBER force field on the same set of three-dimensional structures to check up on any possible force-field dependencies. In agreement with the previous findings, the calculated raw ligand-receptor interaction energies were highly correlated with the inhibitory activities (r2 = 0.81), and the linear regression model relating both magnitudes had an acceptable predictive ability both in internal validation tests (q2 = 0.79, SDEPcv = 0.61) and when applied to the external set of 16 different inhibitors (SDEPex = 1.08). When the interaction energies were further analyzed using the COMBINE formalism, the resulting PLS model showed improved fitting properties (r2 = 0.89) and provided better estimations for the activity of the compounds in the external data set (SDEPex = 0.83). Computation of the electrostatic part of the ligand-receptor interactions by numerically solving the Poisson-Boltzmann equation did not improve the quality of the linear regression model. On the contrary, incorporation of the solvent-screened residue-based electrostatic interactions and two additional descriptors representing the electrostatic energy contributions to the partial desolvation of both the ligands and the receptor resulted in a COMBINE model that achieved a remarkable predictive ability, as assessed by both internal (q2 = 0.73, SDEPcv = 0.69) and external validation tests (SDEPex = 0.59). Finally, when all the inhibitors studied were merged into a single expanded set, a new model was obtained that explained 91% of the variance in biological activity (r2 = 0.91), with very high predictive ability (q2 = 0.81, SDEPcv = 0.66). In addition, the COMBINE analysis provided valuable information about the relative importance of the contributions to the activity of individual residues that can be fruitfully used to design better inhibitors. All in all, COMBINE analysis is validated as a powerful methodology for predicting binding affinities and pharmacological activities of congeneric ligands that bind to a common receptor.