Prediction of DILI in Humans: Investigating Opportunities in DILIst Data, Chemical Features, Machine Learning and Rule-Based Methods

Despite recent scientific developments, combating Drug Induced Liver Injury (DILI) associated risks is still a major challenge and there is an urgent need for breakthroughs in the assessment of DILI. To address this, we have built and evaluated several Machine Learning (ML) and rule-based models. Our approach is focused on three aspects: i) selection of three new high-quality datasets, DILI-876, DILI-955, and DILI-1148 from DILIst (DILI severity and toxicity) dataset, by applying several scientific criteria. The datasets DILI-876, DILI-955, and DILI-1148 contain drugs that fall within three chemical spaces, namely, drugs with molecular weights of ⩽500, ⩽ 750, and "no molecular weight bias" respectively. ii) Evaluation of the influence of three types of data division methods and chemical features on model performance. iii) Evaluation of four ML and one rule-based method for developing reliable DILI models. These efforts resulted in high-performing DILI models that are superior to reported ML models, based on AUC, F1, specificity, and most significantly, with superior performance in the case of DILI positive and negative drugs. The test set characteristics of the best ANN DILI model developed using DILI-876 are AUC-0.908, accuracy-0.804, sensitivity-0.804, specificity-0.815, balanced accuracy-0.809, MCC-0.594, precision-0.821, and F1-0.807. The best SVM DILI models showed similar performance characteristics. The order of predictive performance of the ML models derived from the three datasets is DILI-876>DILI-955>DILI-1148, and the order of performance of the ML methods is ANN~SVM>LR>RF. The order of performance of the models built using the three chemical features is fingerprints>descriptors>fragments. From the performance analysis of our models for individual chemical and DILI classes, we infer that our models performed equally well for DILI negative drugs e.g., the success level of prediction is 81.355% for "vNo" drugs, as reflected by the excellent specificity of the models for the imbalanced DILI data. Similarly, they performed equally well in the case of DILI positives, as reflected by the excellent sensitivity of our models. It is evident from the success level of prediction of "vMost" (81.818%), "vLess" (90.163%), and "others" (73.333%), DILI concerns, that our models are excellent candidates for easy adoption in real-world applications.