Modular fully-connected convolutional neural network: a new method for searching biological active compounds

Relevance: Machine learning methods are widely used today in the search for pharmacological compounds. The nature and internal structure of chemical and biological data are highly specific, and the vast majority of drugs act simultaneously on multiple biotargets. Given this, the development of new artificial neural network architectures for analyzing the relationships between the biological activity and structure of chemical compounds, taking into account the specific nature of chemical and biological information and the interactions of compounds with multiple biotargets, is important and scientifically relevant task.

Objective: To create a new architecture for a modular multi-target fully connected convolutional neural network based on correlation convolution of energy spectra of multiple docking to multiple biotargets, for the in silico searching biological active compounds.

Materials and methods: Ensemble multiple docking of 234 compounds with antimicrobial activity against S. aureus and 537 compounds with anxiolytic activity into 10 and 22 relevant biotargets, respectively, and the generation of their energy spectra of multiple docking were performed using the original MSite program and AutoDock Vina program. Using the obtained energy spectra of multiple docking, two modular multi-target fully-connected convolutional correlation neural networks were constructed using the original FCCorNet program. These networks describe the dependences of the levels of antibacterial activity against S. aureus and anxiolytic activity of chemical compounds on the energies of their modular neural networks. The accuracy and statistical significance of the constructed neural network models were assessed using correlation analysis, one-way analysis of variance, and threshold classification.

Results and discussion: The accuracy of the constructed neural network model for the antimicrobial S. aureus activity was Acc = 78.9 %, with statistical significance p = 3.44 × 10^-12. The accuracy of the constructed neural network model for anxiolytic activity was Acc = 61.3 %, with statistical significance p = 6.68 × 10^-8. The accuracy of predicting the antimicrobial S. aureus activity exceeds the accuracy of predicting the anxiolytic activity, which is probably due to a more complex systemic multi-target mechanism for implementing psychotropic effects, in comparison with the antibacterial action of chemical compounds. The obtained results prove the high validity of using the new architecture of the modular multi-target fully connected convolutional correlation neural network based on the energy spectra of multiple docking for in silico searching biological active substances.

Conclusion: A new artificial intelligence method for in silico searching biological active compounds has been developed: a modular multi-target fully connected convolutional correlation neural network based on the energy spectra of multiple docking into relevant biotargets. Multivariate statistics methods demonstrated high accuracy and statistical significance of the constructed neural network models, reaching p = 3.44 × 10^-12 for antibacterial activity against S. aureus and p = 6.68 × 10^-8 for anxiolytic activity. The developed methodology can be used for in silico searching new highly active compounds with various types of systemic multi-target biological and pharmacological activity, taking into account their integrated affinity for relevant target proteins.