**4. Discussion**

Because of polypharmacy, when several drugs are taken simultaneously, the phenomenon of metabolic DDIs may appear. DDIs manifest in the mutual influence of drugs on their biotransformation, its slowdown, or acceleration, and leads to a change in the pharmacological action of drugs.

To avoid drug withdrawal from the market due to DDIs, pharmaceutical companies perform in vitro and in vivo studies. Physiologically based pharmacokinetic (PBPK) modeling is the in silico method of DDIs prediction that has already proved its applicability in the drug discovery and development process. It is clear that in silico methods will be used more intensively to reduce investigation costs [3].

The main problem we consider is the study and use of the relationship of chemical compound structure and the phenomenon of metabolic DDIs mediated by the seven isoforms of cytochrome P450 most involved in drug metabolism. The models created can be applied for virtual and not-yet-synthesized molecules using only their structural formulas. The implementation of PoSMNA descriptors and the PASS program algorithm for DDIs prediction at the level of cytochromes P450 makes it possible to consider a pair of molecules interacting as one entity without specifying the roles (substrate, inhibitor or inducer, "object" or "precipitant" drug) of particular substances in the DDI process. Such an approach is unique and has already been used to create models for DDIs severity prediction [15,16]. However, when predicting the DDIs severity without taking into account concrete pharmacokinetic or pharmacodynamic DDIs mechanisms, the accuracy of the prediction was not high enough, as compared to that obtained in the current study that considers only pharmacokinetic DDIs mediated by the seven cytochrome P450 isoforms (0.84 for three classes and 0.75 for five classes of severity vs. 0.92 for DDIs prediction mediated by cytochrome P450 isoforms). Such a lower accuracy may be explained by the unclear separation of DDIs of these severity classes among themselves and the cases of DDIs in neighboring classes in the training set and by neglecting the DDIs mechanisms. In this study, the average accuracy of DDIs prediction at the level of cytochrome P450 isoforms is higher (0.92) due to the structural specificity of substances from the pairs that interact at a particular level of the cytochrome P450 isoform. Further research should combine the prediction of DDIs severity at the level of a particular metabolic enzyme. To achieve this goal, it is necessary to expand, improve, and refine the training sets.

**Author Contributions:** Conceptualization, A.V.D.; methodology, D.A.F. and A.V.R., software D.A.F. and A.V.R.; investigation, D.A.K., P.V.P., A.V.R. and A.V.D.; resources, D.A.K., P.V.P., A.V.R. and A.V.D.; data curation, D.A.K., P.V.P., A.V.R. and A.V.D. writing—original draft preparation, A.V.D. and V.V.P.; writing—review and editing, A.V.D. and V.V.P.; supervision, A.V.D., D.A.F. and V.V.P.; project administration, A.V.D., A.A.L., D.A.F. and V.V.P. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research was supported by the Russian Science Foundation (Grant No. 17-75-20250).

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** Data are contained within the article.

**Conflicts of Interest:** The authors declare no conflict of interest.
