The simplest way of searching for a new biologically active substance is to search for a drug that will bind with a particular biological target. However, hits found within such a procedure very often fail at the subsequent stages of the drug design pipeline, due to the adverse effects caused by them, which is most often related to the interaction with the so-called off-targets. Therefore, nowadays, the most advanced procedures consider multiple targets at the same time, following the concept of polypharmacology.

In the study, we developed a polypharmacological machine-learning-based module for virtual screening. It is part of a multi-tool platform, enabling comprehensive approach to solve polypharmacological tasks. The module allows evaluation of compound libraries towards almost 200 protein targets, out of which several (selected by the user) can be consider simultaneously. The Support Vector Machines were trained and optimized for each target separately on the data present in the ChEMBL database. For compound representation, three fingerprints were used: one of a hashed typeand two substructural ones.

We demonstrate several examples of compounds which failed at different stages of the drug discovery process how our tool works and how it can support methods of searching for new drugs. Such optimization procedures are very important from the economical point of view, as they can save time and money devoted to research on compounds, which fail at further stages of their development.

Acknowledgments

This study was supported by grant No. LIDER/37/0137/L-9/17/NCBR/2018 from the Polish National Centre for Research and Development