Development of machine learning-based prediction of drug-induced liver injury using FAERS

<p>【Background and Objectives】Drug-Induced Liver Injury (DILI) is one of the major cause of discontinuation of clinical trials and post-marketing. Many of them are idiosyncratic and difficult to detect in animal and clinical studies. In this study, we used the database of adverse drug event to focus on DILI in clinical settings, and aimed to build a machine learning model to predict the risk of DILI from structural information.</p><p>【Methods】 The FDA Adverse Event Reporting System (FAERS) from the fourth quarter of 1997 to the second quarter of 2020 was used to extract drugs. To remove duplicate reports and to standardize drug names, we used JAPIC AERS, which was cleaned from it by the Japan Pharmaceutical Information Center. The target adverse effect was "drug-related liver injury" in the standard MedDRA ver23.0 search formula. The DILI dataset was created by using the Proportional Reporting Ratios(PRR) method and the number of reports as DILI positive/negative definitions. As explanatory variables, we used 2D descriptors calculated by alvaDesc, which is a molecular descriptor calculator. DILI prediction model was constructed by Random Forest (RF).</p><p>【Results and Discussion】We constructed the prediction model by RF using JAPIC AERS (positive: 446 drugs, negative: 223 drugs), and its ROC-AUC was 0.79. It is suggested that the method of extracting DILI-positive and -negative substances from FAERS and constructing a machine learning model by the chemical structure information can efficiently predict DILI. This method is expected to be used for screening in drug development.</p>