*2.2. PASS*

The PASS software (Laboratory of Structure-Function Based Drug Design, Institute of Biomedical Chemistry, Moscow, Russia) [19] is based on the advanced naïve Bayes classifier and predicts the profiles of biological activity for drug-like compounds. The PASS algorithm creates a classification model of structure–activity relationships based on the training set with structures and known biological activities of known pharmaceutical agents. The PASS prediction results are presented as a ranked list of various biological activities with calculated probabilities *Pa* ("to be active") and *Pi* ("to be inactive"). The most probable activities are those predicted with the maximum value Δ*P* = *Pa* − *Pi*. Currently, PASS predicts more than 8000 types of biological activities, including pharmacological effects, mechanisms of action, influences on gene expression, toxic and adverse effects, and interactions with metabolic enzymes and transporters. Biological activities for particular molecules in the PASS program are represented qualitatively as "active" or "inactive." The structural formulae of drug-like organic compounds are described by Multilevel Neighborhoods of Atoms (MNA) descriptors.

The prediction of DDIs occurring due to interactions with various cytochrome P450 isoforms is similar to the prediction of biological activity using the PASS software. For DDIs prediction mediated by cytochrome P450 isoforms, the input data are represented by the pairs of structural formulas of studied drug-like compounds. The prediction results for each pair of compounds are presented by the probabilities *Pa* and *Pi* lists, which estimate DDIs that may occur due to interactions with CYP1A2, CYP2B6, CYP2C19, CYP2C8, CYP2C9, CYP2D6, and CYP3A4.

#### *2.3. Pairs of Substances Multilevel Neighborhoods of Atoms Descriptors*

To describe the structures of drug pairs, we used PoSMNA descriptors instead of the MNA descriptors applied in the standard PASS software version [19]. PoSMNA descriptors can be used to predict various phenomena, e.g., synergistic effects of two drugs or the prediction of DDIs. Initially, we developed and used PoSMNA descriptors to predict DDIs severity [15,16]. The set of PoSMNA descriptors is the direct product of a combination of two sets of MNA descriptors for each molecule in the DDI pair as {a,b,c, . . . } × {d,e,f, . . . } = {ad,ae,af, ... , bd,be,bf, ... , cd,ce,cf, ... }. MNA/2 (second level of MNA descriptors) for non-hydrogen heavy atoms is used for PoSMNA creation. The MNA descriptors are ordered lexicographically for each pair of compounds, for example, from string "C(C(CCC)C(CC-H)C(CC-H)) C(C(CCC)C(CC-H)O(CC))" to "-O(-C(-C-C-O)) -O(- C(-C-O-O))" (see the examples of PoSMNA descriptors for warfarin and naproxen in Figure 1).

**Figure 1.** Representation of the warfarin and naproxen molecules by Pairs of Substances Multilevel Neighborhoods of Atoms (PoSMNA) descriptors.

To create the models for DDIs prediction, PoSMNA descriptors were generated for all pairs of compounds with known DDIs mediated by CYP1A2, CYP2B6, CYP2C19, CYP2C8, CYP2C9, CYP2D6, or CYP3A4 isoforms of cytochrome P450 in the training set.
