Development of a novel in silico classification model to assess reactive metabolite formation in the cysteine trapping assay and investigation of important substructures
Predicting whether a compound can cause drug-induced liver injury (DILI) is difficult due to the complexity of its mechanism. The production of reactive metabolites is one of the major causes of DILI, particularly idiosyncratic DILI. The cysteine trapping assay is one of the methods to detect reactive metabolites which bind to microsomes covalently. However, it is cumbersome to use 35S isotope-labeled cysteine for this assay. Therefore, we constructed an in silico classification model for predicting a positive/negative outcome in the cysteine trapping assay to accelerate the drug discovery process. In this study, we collected 475 compounds (436 in-house compounds and 39 publicly available drugs). Using a Message Passing Neural Network (MPNN) and Random Forest (RF) with extended connectivity fingerprint (ECFP) 4, we built machine learning models to predict the covalent binding risk of compounds. The 5-fold cross-validation (CV) and hold-out test were evaluated in random- and time-split trials. Additionally, we investigated the substructures that contributed to positive results in the cysteine trapping assay through the framework of the MPNN model. In the random-split dataset, the AUC-ROC of MPNN and RF were 0.698 and 0.811 in the 5-fold CV, and 0.742 and 0.819 in the hold-out test, respectively. In the time-split dataset, AUC-ROC of MPNN and RF were 0.729 and 0.617 in the 5-fold CV, and 0.625 and 0.559 in the hold-out test, restrictively. This result suggests that the MPNN model has a higher predictivity than RF in the time-split dataset. Hence, we conclude that the in silico MPNN classification model for the cysteine trapping assay have better predictive power. Furthermore, most of the substructures that contributed positively to the cysteine trapping assay were consistent with previous reports such as propranolol, verapamil, and imipramine. This is a new machine learning model that can determine the outcome of the cysteine trapping assay, namely accurately predicting the covalent binding risk as the one of factors of idiosyncratic DILI. We believe that this can contribute to mitigating DILI risk for reactive metabolites at the early stages of drug discovery.