Early Detection of Drug Toxicity using Neural Networks

Abstract

Toxicity prediction of drugs is a critical step in the drug development process, as it evaluates the safety of drugs. An abundance of resources go into the development of new drugs, yet only 12% of all drugs are considered by the FDA, since many potential drugs are toxic. By the time toxicity has been identified in conventionally developed drugs, anywhere from $1 to $2 billion could’ve been invested. Therefore, it is imperative to have an early detection of drug toxicity. AI can be used to predict toxicity, by using QSAR and machine learning methods. Some ML based (DNN, SVM, RF) solutions have been proposed that use only molecular features of the compounds, not molecular structure. The goal of this project was to create a GNN ML model (which uses both atomic features and molecular structure information) that could make better toxicity predictions. Tox21 toxicity dataset was used for all training and evaluation. Three GNN models were created and then compared to a SVM model. Statistical analysis of results showed that the GNN models performed better than the SVM model, with the GNN models having better F1-scores (3.34%-6.44% improvement) and MCC values (5.48%-9.40% improvement). Results of this project show that GNN-based models have better toxicity prediction compared to SVM-based models. GNNs have great potential to augment existing QSAR methods used in predicting toxicity in the drug development industry, thus reducing cost, time and resources invested, and alleviating ethical concerns of animals/clinical trials when compared to conventionally developed drugs. This model can help in molecule toxicity prediction for pharmaceutical and other industries.