Predicting Ames Mutagenicity Using Conformal Prediction in the Ames/QSAR International Challenge Project

Area of Science:

• Computational toxicology
• Cheminformatics
• Predictive modeling

Background:

• Ames mutagenicity testing is crucial for drug safety.
• Developing accurate predictive models for mutagenicity is an ongoing challenge.
• Conformal prediction offers a robust framework for uncertainty quantification in predictive models.

Purpose of the Study:

• To develop and validate predictive models for Ames mutagenicity using conformal prediction.
• To assess the impact of data sources (public vs. DGM/NIHS) on model performance.
• To investigate the role of data consistency in achieving reliable mutagenicity predictions.

Main Methods:

• Random Forest models were employed for classification.
• Signature molecular descriptors were used as features.
• Conformal prediction was applied to quantify prediction uncertainty and ensure model validity.
• Models were trained and tested using both public and DGM/NIHS datasets.

Main Results:

• Conformal prediction models demonstrated valid and predictive capabilities for Ames mutagenicity.
• Excluding weakly mutagenic compounds improved model validity for mutagenic compounds when using combined datasets.
• Models trained solely on DGM/NIHS data exhibited superior validity for both mutagenic and non-mutagenic classes compared to models using public data.
• Data consistency significantly impacts predictive quality and model validity.

Conclusions:

• Conformal prediction is effective for building valid Ames mutagenicity classification models.
• Data consistency, particularly using curated datasets like DGM/NIHS, is paramount for reliable toxicological predictions.
• The findings highlight the importance of careful data selection and curation in cheminformatics and computational toxicology.