DeepDILI

DeepDILI predicts drug-induced liver injury (DILI) risk for chemical compounds using deep learning and molecular descriptors to support safety assessment during drug development.

Key Features:

• Integration of Machine Learning and Deep Learning: Combines conventional machine learning algorithms with a deep learning framework and uses model-level representations alongside Mold2 molecular descriptors.
• Performance Evaluation: Evaluated against five conventional machine learning algorithms and two state-of-the-art ensemble methods, achieving a Matthews correlation coefficient (MCC) of 0.331.
• Model-Level vs. Molecule-Based Representation: Model-level representation outperformed molecule-based representations, improving MCC by 25.86% on the test set.
• Explainability and Descriptor Analysis: Identified 21 chemical descriptors enriched in association with DILI outcomes, supporting descriptor-level interpretation of predictions.
• Descriptor Comparison: Compared Mold2, Mol2vec, and MACCS descriptors, with the Mold2-based model exhibiting superior performance.
• Application to Repositioning: Applied to assess DILI concerns for potential COVID-19 treatments derived from drug repositioning candidates.
• Therapeutic Category Discrimination: Showed enhanced discrimination within the World Health Organization therapeutic category 'alimentary tract and metabolism'.

Scientific Applications:

• Preclinical DILI screening: Screening compounds for DILI risk in preclinical drug development using molecular descriptors and model-level representations.
• Prospective safety prediction: Predicting the DILI potential of newly approved drugs based on data from earlier approvals.
• Drug repositioning assessment: Assessing DILI risk for repositioned drug candidates, including potential COVID-19 treatments.
• Molecular interpretation: Identifying chemical descriptors associated with DILI to inform mechanistic interpretation and safer design.

Methodology:

Combines conventional machine learning algorithms with a deep learning framework, employs model-level representation and Mold2 molecular descriptors, compares Mold2 against Mol2vec and MACCS descriptor sets, evaluates performance using Matthews correlation coefficient (MCC) against five conventional ML algorithms and two ensemble methods (MCC = 0.331), reports a 25.86% MCC improvement for model-level representation on the test set, and identifies 21 DILI-associated chemical descriptors.