Discovery of Diphenylsilane Compounds as Potential Inhibitors of Zn-Metalloproteinase Thermolysin Using Artificial Neural Networks.

Introduction/objective: Artificial neural networks are very powerful machine learning and artificial intelligence tools for computer-aided drug discovery. This method offers advantages over traditional approaches related to saving time and money. The aim of this work is to develop machine-learning artificial neural networks for computer-aided discovery of potential thermolysin metalloprotease inhibitor drug candidates.

Methods: In this work, MLP (Multilayer Perceptron) and RBF (Radial Basis Function) neural networks with a general 5:n:1 architecture are developed to find a non-linear correlation between five molecular descriptors obtained by genetic algorithm, and the inhibition of the enzyme thermolysin, expressed as pKi. The AD (applicability domain) of the model was determined using the AMBIT Discovery software, and the in silico activity profile of a series of diphenylsilanes obtained by chemical synthesis was evaluated.

Results: The proposed models show a better fit than the linear model in both series (R2 MLR = 0.71 for the training set and R2 MLR = 0.72 for the prediction set). In the case of the MLP, R2 values close to 0.90 are found and all the compounds are inside the AD of the model. Although the RBF-type models show instability in training, networks with a performance greater than 0.80 are found. However, only the MLP-type models are taken into account to predict the activity of a series of 17 diphenylsilanes. Finally, three compounds are identified as the most promising thermolysin inhibitors.

Conclusion: This methodology offers advantages over traditional methods related to saving time and money. Furthermore, the results obtained suggest that the three identified compounds could be used for the treatment of cardiovascular pathologies because of their homology with human vasopeptidases.