ИСТИНА

Background The idea about correlation between simple molecular properties and compounds' behavior as drugs has gained ground since reporting the Lipinski's Rule of 5. This approach has been reflected in numerous studies as various rigid cut-offs for compounds' properties. With the development of the idea, properties were combined into a single score which reflects the overall efficacy of the compound. In this multi-optimization task one has to choose methods of properties' calculation. Aims In this work we pursue the aim to determine which methods are likely to fit for compounds properties' determination in multi-parameter optimization. Does the result of compound's assessment depend on method's variation? At the same time it is interesting to comprehend if the rank correlation for compounds in a database is the same when the different methods for property prediction are used? Methods The investigation was carried out by StarDrop software. The analysis is based on the idea of properties' desirability. Changes of compound’s descriptors (properties) values are represented as desirability functions, which are integrated into a single score. This score characterizes a compound for a particular task. Quantitative Estimate of Drug-likeness (QED) and two indices characterising the performance of compounds as putative intravenous or oral drugs for the central nervous system were used as combinations of individual desirability functions. Several databases were used: manually curated GSK-3β inhibitors databases and random subsets of ZINC, Commercial Compound Collection, and ZINClick databases. In all cases descriptors were calculated by well-known methods. Results MPO schemes with advanced desirability functions show better performance than schemes based on simple cut-offs or on step functions. Limitations in the use of step functions may be overcome by taking into account the uncertainty of descriptors' calculation. Different calculation schemes for count-based descriptors give different results for the same compounds. Only the schemes used for training give acceptable results in MPO analysis. There is no difference between the current state-of-the-art methods for calculation of continuously distributed properties in the course of MPO. These schemes are completely interchangeable. It was proved that consistent MPO results may be obtained only for compounds belonging to the space of common Ro5-compliant drugs and drug candidates. Property prediction outside of this space should be taken with a great care. Conclusion Multi-parameter analysis of compounds revealed that methods of descriptors' calculation play the critical role in the evaluation of compounds. It is important to understand the scheme on which descriptors are based and use only methods, realization of which is clear and well-defined. Blind use of calculation methods does not guarantee correct results.