Simulation of alternative binding modes in a structure-based QSAR study of HIV-1 protease inhibitors

We have used a published set of inhibitors of HIV-1 protease to build a COMBINE-type structure-based QSAR model with good predictive ability (r2 = 0.90, q2 = 0.69). Since the compounds in the training series exhibit most of their structural variability on one-half of the pseudosymmetrical binding cavity and only one binding orientation was explored for each molecule, the model describes mainly the effect of the structural changes on interactions involving only one-half of the binding cavity (pockets S1' and S2'). Thus, the model cannot be expected to give accurate predictions for new compounds exhibiting structural variation in both halves. The model does in fact show a tendency to underpredict slightly the biological activity of the molecules in the external test set. In an attempt to improve the quality of the model, both possible orientations of the ligands are now considered so that structural variation takes place in all binding pockets. One possibility would have been to build an additional set of complexes with the inhibitors docked in a reversed orientation. The alternative we have explored, however, consists of manipulating the data matrix describing the interaction energies so that each row is duplicated and the order of the variables in the duplicated rows is swapped between subunits. This simple approach has produced a new model that is similar in quality to the original model (r2 = 0.89, q2 = 0.64) but lacks the tendency to underpredict the activity of the compounds in the external set. Moreover, since equivalent residues are assigned equivalent weights, the model is insensitive to ligand orientation and is easier to interpret.