A maximum common subgraph kernel method for predicting the chromosome aberration test

Area of Science:

• Computational chemistry
• Toxicology
• Machine learning

Background:

• The chromosome aberration test is crucial for assessing chemical and drug genotoxicity in mammalian cells.
• Limitations in early drug discovery necessitate accurate predictive models for genotoxicity.
• A need exists for models that guide structure-activity relationship optimization in drug development.

Purpose of the Study:

• To develop a machine learning model for predicting chromosome aberration test outcomes based on chemical structure.
• To provide an interpretable model that identifies structural elements influencing genotoxicity predictions.
• To establish a benchmark for predictive genotoxicity modeling using a high-quality, publicly available dataset.

Main Methods:

• A machine learning approach combining a maximum common subgraph kernel with a support vector machine.
• Utilizing chemical graph representations to measure compound similarity.
• Developing a classification model for predicting genotoxicity assay results.

Main Results:

• The proposed method achieved significantly higher performance compared to standard industry methods and other graph kernel approaches.
• The model provides interpretability, visualizing structural contributions to the prediction.
• A high-quality, curated dataset and cross-validation protocol were established and made public.

Conclusions:

• The developed machine learning model offers a powerful and interpretable tool for predicting chemical genotoxicity.
• This approach can enhance early-stage drug discovery by guiding structure optimization and reducing experimental testing.
• The publicly available dataset and protocol facilitate further research and validation in predictive toxicology.