Abstract
Thrombin plays a central role in thrombosis and hemostasis. Inhibition of thrombin is a prime target for therapeutic intervention of thrombosis. A considerable number of experimental structures of both thrombin and trypsin complexes with their non‐covalently bound inhibitors is available and they offer an excellent database for development of chemometric models for prediction of inhibitors binding affinities towards thrombin and trypsin. Counter‐propagation artificial neural network (CP‐ANN) was applied as a modeling method. As input we have used molecular electrostatic potentials (MEP) of inhibitor–enzyme complexes computed at the contact surface of ligand and protein. The variable selection was performed with the implementation of a genetic algorithm (GA) and the original number of variables (points in which the MEP was computed) have been reduced by more than 70%. In order to assess the influence of non‐covalent interactions of inhibitors with protein environment a second model based on molecular descriptors (MD) for the isolated inhibitors was developed as well and the predictive ability of this approach was compared with the former. All optimized models were validated with compounds from the external validation set. A significantly improved predictive ability in comparison with our previous work was achieved with an RMS error of 0.83 and 0.34 log units for prediction of binding affinity pKi for thrombin and trypsin, respectively, in the external validation compound set. Copyright © 2007 John Wiley & Sons, Ltd.