Scientific relevance. Studies of the toxicological and pharmacokinetic properties of medicinal compounds are a crucial stage of preclinical research; unsatisfactory results may invalidate further drug development. Therefore, the development of in silico methods for a preliminary pre-experimental assessment of toxicological and pharmacokinetic properties is a relevant and crucial task.Aim. The study aimed to review current approaches to in silico prediction of the absorption, distribution, metabolism, excretion, and toxicity (ADMET) parameters of pharmacologically active compounds, in particular, the most important toxicological and pharmacokinetic parameters, and to present the results of the authors’ own research in this area.Discussion. According to the review of models for predicting the toxicological properties of chemical compounds (acute toxicity, carcinogenicity, mutagenicity, genotoxicity, endocrine toxicity, cytotoxicity, cardiotoxicity, hepatotoxicity, and immunotoxicity), the accuracy of predictions ranged from 74.0% to 98.0%. According to the review of models for predicting the pharmacokinetic properties of chemical compounds (gastrointestinal absorption; oral bioavailability; volume of distribution; total, renal, and hepatic clearance; and half-life), the coefficient of determination for the predictions ranged from 0.265 to 0.920. The literature review showed that the most widely used methods for in silico assessment of the ADMET parameters of pharmacologically active compounds included the random forest method and the support vector machines method. The authors compared the literature data with the results they obtained by modelling 12 toxicological and pharmacokinetic properties of chemical compounds using the consensus method in the IT Microcosm system and artificial neural networks. IT Microcosm outperformed the models described in the literature in terms of predicting 2 toxicological properties, including carcinogenicity and blood–brain barrier penetration (the prediction accuracy reached 93.4%). Neural network models were superior in predicting 4 toxicological properties, including acute toxicity, carcinogenicity, genotoxicity, and blood–brain barrier penetration (the prediction accuracy reached 93.8%). In addition, neural network models were better in predicting 3 pharmacokinetic properties, including gastrointestinal absorption, volume of distribution, and hepatic clearance (the coefficient of determination reached 0.825).Conclusions. The data obtained suggest that artificial neural networks are the most promising and practically significant direction for the development of in silico systems for predicting the ADMET characteristics of new medicinal products.

A new series of quinoxaline derivatives, 2a–4b, were synthesized and their anxiolytic potential was evaluated in vivo using elevated plus maze (EPM), open field (OF) and light-dark box (LDB) techniques. According to the results of the EPM, four active compounds were found in 2a, 2b, 2c, 4b. Their anxiolytic properties were confirmed in terms of LDB and the most active was compound 2b. In the OF, only 2c had an influence on the locomotor activity of the rodents. Thus, the most promising substance was determined; this was 2b, which has the structure of 2-(2-{[3-(4-tert-butylphenyl)quinoxaline-2-yl]methyl}-4,5-dimethoxyphenyl)-N-methylethan-1-amine hydrochloride. The obtained data were analyzed with the pharmacophore feature prediction approach, which made it possible to compare the structures of the studied compounds with the reference drug diazepam, and to determine the contribution of pharmacophores to the manifestation of the activity under study. ADMET analysis was carried out for compound 2b and the acute oral toxicity of this substance was also tested in vivo. As a result of the study, a promising compound with a high anxiolytic effect and low level of toxicity 2b was found, which is of interest for further preclinical study of its properties.

Simple and multiple docking into the GABAa receptor of its known agonists has been carried out. For these compounds, the energies of simple docking into the GABA binding site and the energy spectra of multiple docking throughout the entire volume of the GABAa receptor were calculated. A model of the dependence of the level of GABAa agonistic activity of chemical compounds on the energy of their simple docking into the GABA binding site has been formed. A neural network model of the dependence of the level of GABAa agonistic activity on the energy spectrum of multiple docking of chemical compounds has been constructed. It is shown that the predictive accuracy of the neural network model based on multiple docking significantly exceeds the predictive accuracy of the model based on simple docking.

The purpose of the study. Development of a new architecture of a fully connected convolutional neural network based on the correlation convolution of multiple docking energy spectra for in silico search for biologically active compounds. Materials. Verified data on structure and antimicrobial activity against S. aureus of 254 known compounds structurally similar to the studied quinazolinone scaffold; validated 3D model of S. aureus peptide deformylase. Methods. A new developed algorithm for constructing a fully connected convolutional correlation neural network based on multiple docking energy spectra. The MSite program was used to construct spaces for multiple docking on a 3D model of peptide deformylase. The AutoDock Vina program was used to perform ensemble multiple docking of known S. aureus antimicrobial compounds to peptide deformylase, and their multiple docking energy spectra were formed. The FCCorNet program was used to construct a fully connected convolutional correlation neural network on the obtained data and calculate the energy values of this neural network. The AutoDock Vina program was used to perform ensemble simple docking of known S. aureus antimicrobial compounds to a specific binding site of peptide deformylase. The dependences of the expressed antimicrobial activity of chemical compounds on the energies of a fully connected neural network and on the energies of simple docking in the binding site of S. aureus peptide deformylase were studied using the methods of one-factor dispersion, discriminant, cluster and ROC analyses. Results. An algorithm for constructing a fully connected convolutional correlation neural network based on multiple docking energy spectra has been developed. 27 spaces for multiple docking have been constructed on a valid 3D model of S. aureus peptide deformylase. Ensemble multiple docking of 254 known S. aureus antimicrobial compounds into 27 spaces of peptide deformylase has been performed. A fully connected convolutional correlation neural network was constructed, reflecting the dependence of the expressed antimicrobial activity of chemical compounds on the energy of this network. Using four methods of multivariable statistical analysis, a study was conducted of the neural network constructed for peptide deformylase and its accuracy was tested in comparison with the results of using simple docking. Very high statistical reliability p=5.37·10-14 of using the developed methodology for assessing the affinity of ligands to a biotarget was shown. In all respects, the accuracy of searching for active compounds using the energy of a fully connected convolutional correlation neural network based on multiple docking exceeds the accuracy of the model built on the basis of the energy of simple docking into a specific binding site. Conclusions. A new artificial intelligence method was developed for in silico searching biologically active substances based on a fully connected convolutional correlation neural network and multiple docking energy spectra.

Simple and multiple docking into the RAGE receptor of its known inhibitors have been carried out. For these compounds, the energies of simple docking into the specific binding site and the energy spectra of multiple docking throughout the entire volume of the RAGE receptor were calculated. A model of the dependence of the level of RAGE inhibitory activity of chemical compounds on the energy of their simple docking into the specific binding site has been constructed. A neural network model of the dependence of the level of RAGE inhibitory activity on the energy spectrum of multiple docking of chemical compounds has been formed. It is shown that the accuracy of the neural network model based on multiple docking significantly exceeds the accuracy of the model based on simple docking. The found model is used in the directed search for new compounds with RAGE inhibitory activity, which will become the basis for the creation of innovative drugs for the treatment of complications in diabetes mellitus.