Multi-target deep-learning convolutional neural network based on correlation convolution of energies spectra of multiple docking: a new method of machine learning for searching biologically active substances

Machine learning methods are widely used today in the search for biologically active substances. Moreover, chemical and biological data have a highly specific structure, and all medicinal substances act simultaneously on several biotargets. Given this, the development of new methods for constructing deep-learning convolutional neural networks to analyze the relationships between multi-target biological activity and the structure of chemical compounds is a relevant and scientifically important task. Purpose of the work: To create a methodology for building multi-target convolutional neural networks of deep learning based on the correlation convolution of multiple docking energies into relevant biotargets. Materials and methods: Ensemble multiple docking of 537 compounds with anxiolytic activity and 234 compounds with antimicrobial activity against S.aureus into 22 and 10 relevant biotargets respectively, and the subsequent generation of their multiple docking energy spectra were performed using the original MSite program and AutoDock Vina program. Using the original FCCorNet program, correlation convolution of the energy spectra of multiple docking was performed and the energies of fully-connected convolutional neural networks were calculated for the specified compounds. The original computer DeepNets program for constructing deep-learning neural networks was developed in Python using the PyTorch library. Multi-target convolutional neural networks of deep learning were trained on two datasets, including the levels of anxiolytic activity and antimicrobial activity against S. aureus of known compounds and the energies of fully-connected convolutional correlation neural networks, and their accuracy was estimated. Results and discussion: The accuracy of the constructed neural network model for anxiolytic activity was Acc = 68.3 %, with a statistical significance of p = 1.1 × 10-9. The accuracy of the constructed neural network model for antimicrobial activity against S. aureus was Acc = 90.5 %, with a statistical significance of p < 1 × 10-15. The accuracy of predicting antimicrobial activity for S. aureus exceeds that of predicting anxiolytic activity, possibly due to the more complex systemic multi-target mechanism underlying psychotropic effects, compared to the antibacterial action of chemical compounds. The results demonstrate the high validity of a new deep-learning convolutional neural network architecture for in silico searches for biologically active substances. Conclusions: A new multi-target deep-learning convolutional neural network architecture based on correlation convolution of energy spectra of multiple docking into a set of relevant biotargets has been developed. The developed methodology can be used for in silico searches for new high active compounds with various types of multi-target pharmacological activity.