**1. Introduction**

For the treatment of complex disorders, patients often take multiple medications at the same time, which potentially cause drug–drug interactions (DDIs). Usually, DDIs are divided into three types: pharmaceutical, pharmacodynamic, and pharmacokinetic [1]. Pharmaceutical DDIs may appear due to physical or chemical interactions, for example, when drugs are mixed in a syringe before infusion, and such DDIs are rare. Pharmacodynamic DDIs may occur when a pair or more co-administered drugs act on the same physiological system or target. Pharmacokinetic DDIs are very common and occur when one of the drugs ("violator" or "precipitant" drug) affects the absorption, distribution, metabolism, or excretion of another drug ("victim" or "object" drug). Such DDIs provoke an increase or a decrease in the exposure of an object drug and lead to a change in drug pharmacological action. In this study, we focused on the pharmacokinetic DDIs at the metabolism level (biotransformation), the so-called "metabolic DDIs."

The most common drug-metabolizing enzymes (DMEs) in the first phase of xenobiotic metabolism in the human body are several isoforms of the cytochrome P450 superfamily. The U.S. Department of Health and Human Services Food and Drug Administration Center for Drug Evaluation and Research (FDA CDER) requires determining which

**Citation:** Dmitriev, A.V.; Rudik, A.V.; Karasev, D.A.; Pogodin, P.V.; Lagunin, A.A.; Filimonov, D.A.; Poroikov, V.V. In Silico Prediction of Drug–Drug Interactions Mediated by Cytochrome P450 Isoforms. *Pharmaceutics* **2021**, *13*, 538. https://doi.org/10.3390/ pharmaceutics13040538

Academic Editors: Dong Hyun Kim and Sangkyu Lee

Received: 16 February 2021 Accepted: 8 April 2021 Published: 13 April 2021

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

drug-metabolizing enzymes (CYP3A, CYP2D6, CYP2C19, CYP2C9, CYP2C8, CYP2B6, or CYP1A2) metabolize the investigational drug during in vitro studies of metabolic DDIs estimates [2].

In silico methods can help prioritize drug discovery efforts by guiding, but not replacing, in vitro and in vivo experiments. Previously, we presented a comprehensive review of the methods for predicting the DDIs related to the inhibition or induction of DMEs [3]. Most of such in silico methods predict DDIs indirectly. A recently presented machine learning (ML) method used different molecular fingerprints to classify compounds as inhibitors or noninhibitors of five major cytochrome P450 isoenzymes [4]. Ligand-based and structure-based methods dealing with substrates, inhibitors [5,6], and inducers [7] of particular DMEs. Results of prediction could help to determine possible DDIs. However, such conclusions are not sufficiently reliable, as the pairs of substances that are substrates and inhibitors (or inducers) of DMEs may not exhibit DDIs. On the other hand, DDIs have often occurred between substances that could act as substrates, inducers, and inhibitors (that may act by various inhibition mechanisms); for example, this is a widespread case for cytochrome P450 CYP3A4 [8]. At best, a pair of potentially exhibiting DDI substances should be considered together during prediction as the whole entity. However, previously developed ligand-based and structure-based computational methods did not consider two substances in pairs simultaneously. Direct DDIs estimation methods for the pairs of substances include structure resemblance and functional similarities methods and literature-based DDIs prediction methods [9–14]. These methods deal with the pairs of substances but require information about the pharmacokinetics and pharmacodynamics [9,14], interaction profile, target and side-effects [10,13], and the phenotypic, therapeutic, chemical, and genomic properties [11] of substances or medical records [12]. It is clear that for new, not-yet-synthesized, and virtual substances, such information does not exist. The results of predictions of this group of methods [9–14] have often been presented as data sets containing a bulk conglomerate of information about potential DDIs predicted between the existing drugs. Such examples include 430,128 [10], 145,068 [13], and over 250,000 [14] records of unknown potential DDIs in the sets of predicted results. However, this bulk of information concerning drug pairs is provided without assessment of the possibility of DDIs manifestation.

The current study aimed to create the computational structure–activity relationship (SAR) models to predict metabolic DDIs mediated by CYP1A2, CYP2B6, CYP2C19, CYP2C8, CYP2C9, CYP2D6, or CYP3A4. We have previously developed models for DDIs severity prediction [15,16] that used the PASS (Prediction of Activity Spectra for Substances) program and PoSMNA (Pairs of Substances Multilevel Neighborhoods of Atoms) substructural descriptors. These models were able to predict the classes of DDIs severity for pairs of molecules according to OpeRational ClassificAtion (ORCA). In the current study, we used the same methods and descriptors but implemented them to predict whether two molecules would manifest metabolic DDIs mediated by the seven cytochromes mentioned above. Due to the limited possibilities of creating an appropriate training set, the stereochemical features of molecules were not taken into account by our descriptors. In addition, in the current realization of the method, DDI predictions were obtained in qualitative mode ("YES" or "NO"). Unlike other ligand-based and structure-based methods [4–8], our approach operated with two substances in pairs at once. This is reasonable for the DDI phenomenon, in which two substances interact simultaneously. It gives a direct indication of DDIs for the pairs of molecules without suggestions of the role of particular compounds, which is not always obvious (without consideration of inhibition or induction of a particular enzyme). In contrast to structure resemblance, functional similarities, and literature-based DDIs prediction methods [9–14], our prediction method uses only structural formulas of compounds; it does not require any information about their biological activity. This means that our method can be applied for not-yet-investigated, new, and virtual substances. Moreover, our method provides a probabilistic assessment of possible DDIs and evaluates the possibility of DDIs manifestation for predicted pairs.

#### **2. Materials and Methods**

*2.1. Information on DDIs and Training Set Creation*

We used DDIs data collected from two sources of information. The first source was DrugBank Version 4.1 (University of Alberta and The Metabolomics Innovation Centre, Edmonton, AB, Canada) [17] that contains information about interactions derived from public drug databases. The second source of DDIs data was the Fujitsu ADME Database (Chemistry & Life Science Group, Fujitsu, Tokyo, Japan) [18].

The final data set includes information from both sources. It was used to create the training set containing information about 2345 pairs of single-component organic compounds that interacted due to CYP1A2, CYP2B6, CYP2C19, CYP2C8, CYP2C9, CYP2D6, or CYP3A4. The detailed information is presented in Table 1.

**Table 1.** The number of drug–drug interactions (DDIs) mediated by various isoforms of cytochrome P450 in the training set.


It is well known that CYP3A4 is the major isoform of human cytochrome P450 involved in drug metabolism and pharmacokinetic DDIs. As we can see from Table 1, the number of DDIs associated with CYP3A4 in the training set is twice as high as the number of pairs for the remaining six cytochromes. It fully reflects the real situation and illustrates that the training set is representative.
