Monte Carlo simulations have been performed on a series of 20 active-site-directed thrombin inhibitors to determine the interactions and energetics associated with the binding of these compounds. Physicochemical descriptors of potential value in the prediction of binding affinities were averaged during simulations of each inhibitor unbound in water and bound to thrombin. Regression equations based on 3-5 descriptors are able to reproduce the experimental binding affinities, which cover a 7 kcal/mol range, with rms errors of 1.0-1.3 kcal/mol, and yield correlation coefficients, r(2), of 0.7-0.8. On the basis of these results, the quantities most important in determining the binding affinities are: (1) the enhancement of van der Waals interactions in going from solution to the bound state, (2) the intramolecular strain induced in the inhibitor upon binding, (3) the number of hydrogen bonds lost in the binding process, and (4) the number of rotatable bonds in the inhibitor. The descriptors are physically reasonable and, in combination with the insights gained from analysis of the simulation structures, suggest directions for the development of improved thrombin inhibitors.