ИСТИНА

Nowadays, along with the classical approach to target-selective drug design, the methods of rational polypharmacology are on the rise. They aim to develop the compounds acting on several targets or pathways simultaneously. A popular solution in the design of both selective and multitarget drugs is based on virtual screening of large compound libraries. In this approach, the quality of a scoring function used in the screening process plays a critical role. Most of the scoring functions currently used in molecular docking are of general nature, attempting to provide binding energy estimates across a wide range of targets. However, their accuracy is often limited, and this can cause problems in the design of multitarget drugs that involves virtual screening against several targets, leading to the multiplication of uncertainty. Thus, in order to differentiate between decoys and actives, it is necessary to select a suitable scoring function for each target or to construct scoring functions specifically for particular targets. We have developed an approach to constructing such functions using the machine learning methods. The empirical potentials calculated using the AutoDock Vina docking software are used as descriptors. The models were built using the RF, SVM, LDA, ANN and kNN machine learning techniques. This approach was successfully used to design multitarget inhibitors of Tankyrase and PI3Kα enzymes that have good potential for the development of drugs against colon cancer. The results based on the classification models are also compared to the application of the built-in scoring functions in AutoDock Vina software.