Monte Carlo statistical mechanics simulations were used in combination with the extended linear response (ELR) approach to develop a model to predict the activities of kinase inhibitors. One hundred forty eight inhibitors of three protein kinases, cyclin-dependent kinase 2 (CDK2), lymphocyte-specific kinase (Lck), and p38 mitogen-activated protein kinase were considered. The inhibitor sets for the individual kinases were analyzed first, and ELR models using only three descriptors were obtained with correlation coefficients, r(2), of 0.7-0.8. Models for each pair of kinases were then developed and used to predict the activities of the inhibitors for the remaining kinase with resultant q(2) values of 0.71 (CDK2), 0.70 (Lck), and 0.54 (p38). Finally, the three datasets were combined to yield a general ELR model for kinase inhibition; with just three physically reasonable descriptors, EXX, DeltaHB(total), and DeltaSASA, the r(2) and leave-one-out q(2) are 0.69 and 0.67. The optimization of the model was confirmed using a genetic algorithm. The descriptors reflect the structural requirements for strong inhibition: good steric and electrostatic complementarities between inhibitor and protein, limited loss of hydrogen bonds for the inhibitor upon binding, and increased burial of surface area of the inhibitor.