ИСТИНА

Structure-based virtual screening using molecular docking has become a standard procedure even for large libraries of commercially available compounds such as ZINC, and its performance is critical for the search for new drugs. It directly depends on the ability of the docking scoring functions to distinguish between the active and inactive compounds. We present a method for creating target- or family-specific scoring functions that are commonly superior to the general purpose scoring functions in terms of their discriminative power. They were built by various machine learning methods (Random Forest (RF), k Nearest Neighbors (kNN), Support Vector Machines (SVM), Linear Discriminant Analysis (LDA) and Deep Neural Networks (DNN). The easily interpretable empirical interaction potentials from the AutoDock Vina / Smina scoring function were used as descriptors in these models. The proposed approach has been applied to create specialized scoring functions for the inhibitors of the Tankarase and PI3K enzymes that could be potential antitumor targets. The best results were obtained for the models based on the Deep Neural Networks (DNN) and the Linear Discriminant Analysis (LDA) approaches. In various external tests, the target-specific functions outperform the “Vina” and “Vinardo” scoring functions. The AUC ROC value is increased from the 0.82 for “Vinardo” to 0.90 for the DNN. The difference in computational complexity compared to the standard approaches is insignificant, allowing one to easily employ this method in the virtual screening of large chemical libraries.