X-DILiver: an ensemble learning framework for predicting drug-induced liver injury

Area of Science:

• Pharmacology
• Computational Chemistry
• Toxicology

Background:

• Drug-induced liver injury (DILI) is a major reason for drug withdrawal, impacting over 1000 medications.
• Predicting DILI risk early in drug development is crucial for drug safety and efficiency.

Purpose of the Study:

• To develop a high-accuracy computational method for predicting DILI risk.
• To create a robust predictive framework utilizing the largest available DILI-annotated dataset.

Main Methods:

• Developed X-DILiver, a predictive framework using data augmentation to handle class imbalance and improve robustness.
• Constructed a library of 312 machine learning models with diverse algorithms and molecular features.
• Applied an optimized ensemble strategy, combining extreme gradient boosting and recurrent neural networks.

Main Results:

• The X-DILiver ensemble model achieved an accuracy of 0.64 and MCC of 0.33 on a Tanimoto-filtered external test set.
• The model achieved an accuracy of 0.64 and MCC of 0.38 on a scaffold-unique external test set, outperforming existing DILI prediction models.
• Performance highlights ongoing challenges in DILI prediction despite significant advancements.

Conclusions:

• X-DILiver offers a valuable advancement for DILI risk assessment in drug discovery.
• The tool serves as a reliable method to predict DILI potential, accelerating safer drug development.
• X-DILiver is publicly accessible at http://xdiliver.lile.bio to promote broad adoption.