A drug-induced liver injury prediction model using transcriptional response data with graph neural network

Drug-Induced Liver Injury (Dili) is a major cause of failed drug candidates in clinical trials and withdrawal of approved drugs from the market. Therefore, machine learning-based Dili prediction can be key in increasing the success rate of drug discovery because drug candidates that are predicted to potentially induce liver injury can be rejected before clinical trials. However, existing Dili prediction models mainly focus on the chemical structures of drugs. Since we cannot determine whether a drug will cause liver injury based solely on its structure, Dili prediction based on the transcriptional effect of a drug on a cell is necessary. In this paper, we propose GLIT which is a model that uses transcriptional response data and chemical structures and can be used for drug-induced liver injury prediction. GLIT learns the embedding vectors of drug structures and drug-induced gene expression profiles using graph attention networks in a biological knowledge graph for predicting Dili. GLIT outperformed a baseline model that uses only drug structure information by 7% and 19.2% in terms of correct classification rate (CCR) and Matthews correlation coefficient (MCC), respectively. In addition, we conducted a literature survey to confirm whether the class labels of drugs, in the unknown Dili class, predicted by GLIT are correct.