Machine Learning-Based Models to Predict Drug-Induced Liver Injury (DILI) to Assist Medicinal Chemistry

Clinical Snapshot

Key Findings

• P Value: Not reported — no inferential statistics are generated by this review

• Effect Size: Not applicable — no pooled effect size reported. Individual model performance metrics are discussed qualitatively across studies; the review notes that traditional preclinical methods fail to detect 40–45% of clinical hepatotoxicity cases, establishing the benchmark gap that computational models aim to address

• Primary Outcome: Narrative synthesis of the predictive performance of ML/DL-based DILI prediction models across available datasets and computational approaches; no pooled primary outcome is calculated

• Nnt Or Sensitivity: Not calculated in this review. Individual ML/DL models discussed in the literature show variable sensitivity and specificity for DILI classification depending on dataset, endpoint definition, and chemical space; no pooled sensitivity/specificity estimates are derived. The review highlights that model performance is highly dataset-dependent and that prospective external validation data remain limited

• Confidence Interval: Not reported — no meta-analytic confidence intervals are calculated in this perspective review

Clinical Application

Computational DILI prediction tools are increasingly accessible via open-source platforms and commercial software, making integration into drug discovery workflows feasible for organisations with cheminformatics capability. However, prospective validation in real-world pipelines remains limited, and standardisation of datasets, endpoints, and model benchmarking is still evolving. Clinical translation requires integration with causality assessment frameworks (e.g., RUCAM) and regulatory acceptance pathways, which are not yet established for ML-based DILI predictions. In the Australian context, DILI represents a significant concern for the Therapeutic Goods Administration (TGA), which evaluates hepatotoxicity data as part of new drug registration and post-market surveillance. The TGA's pharmacovigilance framework (including the Database of Adverse Event Notifications, DAEN) captures post-market DILI signals but does not currently incorporate ML-based predictive tools in its regulatory assessment process. Australian pharmaceutical companies and CROs engaged in early drug discovery could benefit from integrating validated ML/DL DILI prediction tools into their preclinical pipelines, potentially reducing late-stage attrition. The RACGP and GESA (Gastroenterological Society of Australia) guidelines on drug-induced liver disease focus on clinical recognition and management rather than predictive modelling. Pharmacogenomic considerations discussed in this review are relevant to Australia's diverse population, including Indigenous Australians and communities with distinct CYP450 polymorphism profiles, though this is not addressed in the paper. PBS implications are indirect — improved early DILI prediction could reduce the cost burden of late-stage drug failures and post-market withdrawals affecting PBS-listed medicines. Primarily applicable to medicinal chemists and drug discovery scientists evaluating hepatotoxicity risk of novel chemical entities during lead optimisation and candidate selection phases. Secondarily relevant to clinical pharmacologists, regulatory scientists, and hepatologists assessing computational DILI risk signals for investigational or approved drugs. Not directly applicable to individual patient management decisions at this stage of evidence maturity.

Abstract

Drug-induced liver injury (DILI) is a leading cause of drug failure and post-market withdrawals. Traditional preclinical methods fail to detect up to 40-45% of clinical hepatotoxicity cases. Computational approaches, particularly those based on machine learning and deep learning (DL), are emerging as promising tools to support medicinal chemistry and early drug discovery, though their predictive capabilities remain under active investigation. In this perspective, we review the development of DILI annotation data sets, tracing their growth from small collections to large, comprehensive resources. We also outline the evolution of computational methods, from simple descriptor-based models to advanced DL and ensemble approaches that incorporate interpretable features. Finally, we highlight recent efforts to integrate standardized causality frameworks, pharmacogenomics, and mechanistic models, aiming to connect computational advances with clinical relevance. This perspective provides valuable insight for researchers and promotes the development of more robust and consensual DILI prediction strategies.

References

1. 1.Evangelista, D., Nelson, E., Tehan, B., Bagnolini, G., Roberti, M., & Bottegoni, G. (2026). Machine learning-based models to predict drug-induced liver injury (DILI) to assist medicinal chemistry. Journal of Medicinal Chemistry. Advance online publication. https://doi.org/10.1021/acs.jmedchem.5c02888