Enhancing DILI toxicity prediction through integrated graph attention (GATNN) and dense neural networks (DNN)

Drug-induced liver injury (DILI) toxicity is a condition when drugs have a destructive effect on the liver organ. The prediction of this toxicity becomes crucial in the drug development process to guarantee that drugs are safe from toxicity. Assessment is usually carried out in the conventional laboratory, which causes a high cost in materials and time. To help solve the problem, computational technology is used to predict DILI toxicity in compounds and drugs. Many researchers have developed the model by using various molecular datasets. The Simplified molecular input line entry system (SMILES) code data was used to represent drugs or compounds. In this work, we proposed the modified model using the reliable dataset from the previous work. We reproduced the best previous model and combined it with the graph attention neural network. After running the proposed model, the performances outperformed almost all performance metrics of the previous model by 75.14 % precision, 85.2 % sensitivity, 39.9 % MCC value, 75.7 % AUC value and 82.5 % F1 score. All the source code used in the experiment can be accessed freely at this link: https://github.com/asw1982/GATNN DNN DILI Toxicity.