DeepDILI: Deep Learning-Powered Drug-Induced Liver Injury Prediction Using Model-Level Representation

Area of Science:

• Pharmacology and Toxicology
• Computational Chemistry
• Machine Learning in Drug Discovery

Background:

• Drug-induced liver injury (DILI) is a primary reason for drug withdrawal.
• Early identification of DILI potential is crucial for drug development safety.

Purpose of the Study:

• To develop and evaluate a deep learning-powered DILI prediction model (DeepDILI).
• To assess the predictive performance of DeepDILI using historical drug data.
• To compare DeepDILI's model-level representation against molecule-based approaches.

Main Methods:

• Developed DeepDILI by integrating conventional machine learning (ML) model-level representations with a deep learning framework using Mold2 descriptors.
• Trained DeepDILI on drugs approved before 1997 to predict DILI potential for later-approved drugs.
• Evaluated performance using Matthews correlation coefficient (MCC) and compared against ML algorithms and deep neural networks.

Main Results:

• DeepDILI achieved a MCC of 0.331, outperforming conventional ML and ensemble methods.
• Model-level representation in DeepDILI significantly improved DILI prediction accuracy (25.86% MCC increase) over molecule-based deep neural networks.
• Identified 21 DILI-associated chemical descriptors and demonstrated superior performance for 'alimentary tract and metabolism' drugs.

Conclusions:

• The DeepDILI model is a promising tool for preclinical DILI risk screening.
• Model-level representations offer advantages over molecule-based approaches for DILI prediction.
• DeepDILI provides enhanced model explainability and has been applied to predict DILI concerns for potential COVID-19 treatments.