Modelling of Drug-Induced Liver Injury with Multiple Machine Learning Algorithms

Motivation: Drug-induced liver injury (DILI) is one of the primary problems in drug development. The current study aims to build predictive models using the physical and chemical properties of compounds causing DILI. Early prediction of DILI may result in redirecting drug development towards compounds with a lower risk of DILI and thus significantly reduce the risk of market failure. Methods: Research was performed on the FDA DILI Rank data set. All compounds were labelled in three DILI classification schemes: FDA DILI concern classification, DILI severity score, and commercial status of the drug. We used five classifiers to build cross-validated predictive models for different binary splits of drugs into high- and low-risk classes of DILI. Results: The best models obtained (AUC = 0.81, MCC = 0.46) discern between harmless and DILI-causing compounds, regardless of DILI severity. The models for other splits of compounds along DILI scales are worse, showing that prediction of the severity of DILI is more difficult. The filter based on the classifier can be used to assess lead compounds in drug development to decrease DILI risk in new drugs.