Accepted_test

Proteochemometrics (PCM) methods expand the applicability domain of structure-activity models compared with the (Q)SAR approach, as such models use data on both the ligand and protein targets. Therefore, it provides the possibility to predict activity spectra for proteins with an unknown spectrum of ligands. PCM appears to be promising in the search for effective drugs against new viral infections. We developed an original PCM method based on the combined description of protein-ligand pairs and tested it on data representing viral proteases and their inhibitors. The description of protein targets was derived from a sequence alignment based on the pairwise 3D comparisons. The multilevel neighborhood descriptors (MNA descriptors) were used to represent ligands. Testing a ligand set in (Q)SAR mode showed a prediction accuracy of about 0.95 in leave-one-out cross validation. PCM models built using the combined descriptors were tested using a procedure involving the exclusion of protein-ligand pairs. The accuracy values obtained in PCM mode by leave pair-out cross validation reached 0.89. The PCM is characterized in decrease of accuracy compared to (Q)SAR, but provides effective prediction in the absence of reliable data on the interaction of small molecules with the target proteins of the corresponding viruses.