Multi-target neural network model of anxiolytic activity of chemical compounds using correlation convolution of multiple docking energy spectra.

Anxiety disorders are one of the most common mental health pathologies in the world. They require searc h and development of novel effective pharmacologically active substances. Thus, the development of new approaches to the search for anxiolytic substances by artificial intelligence methods is an important area of modern bioinformatics and pharmacology. In this work, a multi-target model of the dependence of the anxiolytic activity of chemical compounds on their integral affinity to relevant target proteins based on the correlation convolution of multiple docking energy spectra has been constructed using the method of artificial neural networks. The training set of the structure and activity of 537 known anxiolytic substances was formed on the basis of the previously created database, and optimized 3D models of these compounds were built. 22 biotargets presumably relevant to anxiolytic activity were identified and their valid 3D models were found. For each biotarget, 27 multiple docking spaces have been formed throughout its entire volume. Multiple ensemble molecular docking of 537 known anxiolytic compounds into all spaces of relevant target proteins has been performed. The correlation convolution of the calculated energy spectra of multiple docking was carried out. Using seven training options based on artificial multilayer perceptron neural networks, the multi-target model of depending anxiolytic activity chemical compounds on 22 parameters of the correlation convolution of the multiple docking spectra energy was constructed. The predictive ability of the created model was characterized Acc = 91.2% and AUCROC = 94.4%, with statistical significance of p < 1×10⁻¹⁵. The found model is currently used in the search for new substances with high anxiolytic activity.