A deep neural network for the identification of lead molecules in antibiotics discovery
Michael Idowu Oladunjoye, Olumide Olayinka Obe, Olufunso Dayo Alowolodu · 2022
In this study, we develop a deep neural network (DNN) model, multi-layer perceptron (MLP) to classify the molecules into "active" and "inactive" compounds using a ligand-based virtual screening approach for the lead compounds identification at the early stage of the antibiotic discovery. Lead identification as a major part of virtual screening in the drug discovery process is mostly performed by the quantitative structure-activity relationship (QSAR)-based method. The purpose of applying an artificial intelligence (AI) method is to reduce the time and subsequently the costs that are always associated with the process. The MLP model has several stacks of hidden layers and it used a back-propagation algorithm for the training. The dataset of experimentally known bioactivities of the drug-like compounds and their respective target was obtained from ChEMBL database. A biological target of an antibiotic, dihydrofolate reductase (DHFR), was searched from the database to get its inhibitors' chemical properties and the IC50 values on which the classification was based. One set of the dataset was preprocessed and split into two for the training and validating sets of 80% and 20% respectively. With this approach, the compounds were successfully classified into the desired categories and an accuracy of 0.74 was achieved.