In Silico Search for Antimicrobial Compounds Using Multi-Target Modular Fully Connected Convolutional Neural Network Based on Multiple Docking

Abstract— In ten relevant Staphylococcus aureus targets, multiple docking of 254 known antimicrobial quinazolinone derivatives and 71 new quinazolinone derivatives was performed. Using a new original architecture of a modular multi-target fully connected convolutional neural network based on the correlation convolution of multiple docking energy spectra into relevant targets, a model of anti-S. aureus activity of chemical compounds has been constructed. The threshold value of the total energy of this neural network, separating the high-affinity and low-affinity compounds, has been determined as V0 > 266.5. The limit value of anti-S. aureus activity, the minimum inhibitory concentration, which separated fairly active and low-active compounds, was determined as MIC < 112.5 μg/mL. Using the obtained model, the accuracy of the prediction on the training set for anti-S. aureus activity of quinazolinone derivatives was assessed (Acc = 78.9%), which corresponded to the significance level p = 3.44 × 10–12. Prognosis of the anti-S. aureus activity of 71 new quinazolinone derivatives was carried out, with the forecast efficiency of PreHigh = 73.2%. A substance with anti-S. aureus activity experimentally confirmed in vitro was found. The developed neural network technology can be recommended as a new method of artificial intelligence for in silico searching the antimicrobial substances.