Revealing cytotoxic substructures in molecules using deep learning.

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and development
• Toxicology and risk assessment

Background:

• Late-stage toxicity is a major cause of drug development failure, increasing costs and timelines.
• In silico methods are crucial for early-stage prediction to reduce attrition rates and animal testing.
• Machine learning, particularly neural networks, shows promise in predictive toxicology due to increasing data availability.

Purpose of the Study:

• To investigate a deep learning approach for predicting compound cytotoxicity in early drug discovery.
• To enhance model interpretability by identifying toxicophores responsible for cytotoxic effects.
• To introduce cytotoxicity maps for visual interpretation of structural toxicity relevance.

Main Methods:

• A deep learning model was trained on an in-house dataset of over 34,000 compounds.
• The model achieved a balanced accuracy exceeding 70% in cytotoxicity prediction.
• Deep Taylor Decomposition was employed to identify toxicophores and interpret model predictions.

Main Results:

• The deep learning model demonstrated comparable performance to Random Forest methods.
• Toxicophores, key structural contributors to cytotoxicity, were successfully identified.
• Cytotoxicity maps were generated, providing visual insights into structure-activity relationships for toxicity.

Conclusions:

• Deep learning models can effectively predict compound cytotoxicity.
• The developed interpretability methods (Deep Taylor Decomposition, cytotoxicity maps) offer mechanistic insights into toxic effects.
• This approach can aid in de-risking and optimizing drug candidates during development.