Knowledge-guided deep learning models of drug toxicity improve interpretation.

Area of Science:

• Computational toxicology
• Pharmacology
• Artificial intelligence in drug discovery

Background:

• Drug development faces high attrition rates due to poor understanding of cellular mechanisms behind drug toxicity.
• Conventional classification models for toxicity prediction are often "black boxes," limiting insight into specific toxicity pathways.

Purpose of the Study:

• To develop DTox, an interpretable deep learning framework for predicting compound toxicity and inferring underlying cellular mechanisms.
• To enhance the understanding of drug-induced toxicity pathways in silico.

Main Methods:

• DTox utilizes knowledge-guided neural networks to predict compound responses in toxicity assays.
• The framework infers toxicity pathways for individual compounds, offering interpretability beyond predictive performance.
• DTox was applied to rediscover known mechanisms and differentiate cytotoxicity pathways in HepG2 cells.

Main Results:

• DTox achieves predictive performance comparable to conventional models but with significantly improved interpretability.
• The framework successfully elucidated mechanisms of transcription activation and cellular activities induced by specific drug classes (aromatase inhibitors, PXR agonists).
• Virtual screening with DTox identified compounds with predicted cytotoxicity as higher risk for clinical hepatic phenotypes.

Conclusions:

• DTox provides a powerful, interpretable framework for deciphering cellular mechanisms of drug toxicity computationally.
• This approach can aid in identifying potential safety liabilities early in drug development, reducing attrition rates.
• DTox facilitates a deeper understanding of drug-target interactions and off-target effects related to toxicity.