A Binary Classification Model for Toxicity Prediction in Drug Design

Abstract

Toxicity in drug design is a very important step prior to human or animal evaluation phases. Establishing drug toxicity involves the modification or redesign of the drug into an analog to suppress or reduce the toxicity. In this work, two different deep neural networks architectures and a proposed model to classify drug toxicity were evaluated. Three datasets of molecular descriptors were build based on SMILES from the Tox21 database and the AhR protein to test the accuracy prediction of the models. All models were tested with different sets of hyperparameters. The proposed model showed higher accuracy and lower loss compared to the other architectures. The number of descriptors played a key roll in the accuracy of the proposed model along with the Adam optimizer.