DeepAmes: A deep learning-powered Ames test predictive model with potential for regulatory application.

Area of Science:

• Toxicology
• Computational Chemistry
• Regulatory Science

Background:

• The Ames assay is a global regulatory requirement for assessing mutagenic potential.
• In silico methods, including machine learning, are increasingly used to predict Ames test outcomes.
• Existing predictive models require further enhancement for robust regulatory application.

Purpose of the Study:

• To develop and validate DeepAmes, a novel deep learning model for predicting Ames test results.
• To evaluate DeepAmes' performance against standard machine learning methods.
• To assess the utility of DeepAmes in regulatory science, including its applicability domain.

Main Methods:

• Development of a deep learning model (DeepAmes) using a large Ames dataset (>10,000 compounds).
• Comparison of DeepAmes with five standard machine learning methods.
• Evaluation of model performance, stability, and applicability domain.

Main Results:

• DeepAmes demonstrated superior performance in predicting Ames assay outcomes on a test set of 1,543 compounds.
• The model maintained stable performance even with compounds outside its applicability domain (>30%).
• A revised DeepAmes version showed significantly improved sensitivity (0.87 from 0.47) for regulatory use.

Conclusions:

• DeepAmes provides a high-performance, deep learning-powered predictive model for Ames test results.
• The model's defined applicability domain and performance characteristics support its potential for regulatory applications.
• DeepAmes offers a valuable tool for in silico mutagenicity assessment in consumer product safety evaluations.