ИСТИНА

Virtual screening of compound libraries is an important stage of drug discovery. The selection and ranking of the compounds is based on the estimation of their relevant properties analyzed by the multiparameter optimization (MPO) methods. One of common MPO techniques is based on the calculation of scores integrating the desirability of various properties. In the present work we investigated to what extent the use of different methods of calculation of compound properties (or descriptors) is significant in the course of MPO. We analysed several scoring profiles available on the StarDrop platform: (1) Quantitative Estimate of Drug-likeness (QED), and two indices characterizing the performance of compounds as (2) oral drugs acting on the central nervous system (OCNS) and (3) intravenous drugs not entering into the central nervous system (INCNS). Score values were determined based on the descriptors calculated by different methods. Besides the descriptor calculation schemes available in StarDrop, several other schemes were employed, including ALOGPS, XLogP, our in-house techniques, and methods implemented in ChemAxon and Drug-Like Tool. Calculation schemes were varied for LogP, water solubility (LogS), blood-brain barrier penetration (LogBB), hERG inhibition (hERG), molecular topological polar surface area (TPSA), and number of hydrogen bond acceptors (nHBA). The analysis was performed for several databases: manually curated set of GSK-3β inhibitors, random subsets of ZINC profiled by drug likeness or non-likeness, Commercial Compound Collection, and ZINClick. Pairwise comparison of all the scores and analysis of the ranking stability were performed. The scoring profiles using smooth distribution based desirability functions provided more stable rankings compared to the schemes employing simple cut-offs or step functions. Step-based desirability functions may be used only when the uncertainty of descriptor calculation scheme is known and taken into account. There is no difference in the tested MPO schemes between the current state-of-the-art methods for calculation of continuously distributed properties, such as LogP or LogS. For the count-based descriptors (e.g., nHBA) only the schemes used during the development of a certain MPO scheme lead to acceptable results, because different descriptor calculation schemes give different values for the same compounds. Finally, we have shown that consistent MPO results may be obtained only for compounds belonging to the well-studied chemical space, though not necessarily the drug-like one. Property prediction outside of this space should be taken with a great care.