Skip to main content
CORESTA Meeting, Smoke Science/Product Technology, 2023, Cancun, STPOST 09

Examination of automated growth inhibition classification in Ames test by machine learning

KUM R.; KAIYA K.; ITO H.
Japan Tobacco Inc., Scientific Product Assessment Center, Yokohama, Kanagawa, Japan

The Ames test is a commonly used bacterial bioassay to assess the mutagenicity of chemicals and has been applied to evaluate the mutagenicity of tobacco products. In Ames test, a statistically significant dose-related increase in revertant colonies suggests the mutagenicity of the test substance. However, when the test substance also induces severe cytotoxicity, bacterial growth may be inhibited, and the mutagenicity may not be assessed correctly. Therefore, Organization for Economic Co-operation and Development (OECD) guideline 471 requires information on growth inhibition (GI) as well as the number of revertant colonies. GI is determined based on a reduction in the number of microcolonies (background lawn) and is observed microscopically. This process places a high burden on researchers, and periodic checks are required to prevent differences in classification criteria between individual researchers. Hence, it is desirable to improve the efficiency of the process and to ensure objectivity in the classification of GI.

In this study, we employed machine learning for the classification of five types of GI. Salmonella Typhimurium TA100 was used without metabolic activation, and 1,856 images of GI were obtained to build models to classify types of GI using DataRobot, an automated machine learning platform providing more than 40 models including Auto-tuned K-Nearest Neighbors Regressor, Keras Deep Residual Neural Network Regressor, and Ridge Regressor. Among the provided models, we selected Auto-tuned K-Nearest Neighbors Regressor as the best model based on root mean square error (= 0.1458), and employed this model for prediction of the GI type of images obtained in three independent trials of Ames test. This model achieved average accuracy of 97.6 % for GI level classification of all data in approximately 10 minutes. This approach has the potential to improve the efficiency and objectivity of GI classification in Ames test.