Results of Monte Carlo (MC) simulations for more than 200 nonnucleoside inhibitors of HIV-1 reverse transcriptase (NNRTIs) representing eight diverse chemotypes have been correlated with their anti-HIV activities in an effort to establish simulation protocols and methods that can be used in the development of more effective drugs. Each inhibitor was modeled in a complex with the protein and by itself in water, and potentially useful descriptors of binding affinity were collected during the MC simulations. A viable regression equation was obtained for each data set using an extended linear response approach, which yielded r(2) values between 0.54 and 0.85 and an average unsigned error of only 0.50 kcal/mol. The most common descriptors confirm that a good geometrical match between the inhibitor and the protein is important and that the net loss of hydrogen bonds with the inhibitor upon binding is unfavorable. Other physically reasonable descriptors of binding are needed on a chemotype case-by-case basis. By including descriptors in common from the individual fits, combination regressions that include multiple data sets were also developed. This procedure led to a refined "master" regression for 210 NNRTIs with an r(2) of 0.60 and a cross-validated q(2) of 0.55. The computed activities show an rms error of 0.86 kcal/mol in comparison with experiment and an average unsigned error of 0.69 kcal/mol. Encouraging results were obtained for the predictions of 27 NNRTIs, representing a new chemotype not included in the development of the regression model. Predictions for this test set using the master regression yielded a q(2) value of 0.51 and an average unsigned error of 0.67 kcal/mol. Finally, additional regression analysis reveals that use of ligand-only descriptors leads to models with much diminished predictive ability.