Abstract
The Ames test is a typical test that evaluates the effects (mutagenicity) of chemical substances contained in pharmaceuticals and industrial products on DNA. The test evaluates the mutagenicity of both the chemical itself and metabolites of the compound using rat metabolic enzymes. To evaluate toxicity concerns quickly and inexpensively, in silico (QSAR) technology that predicts toxicity from compound structures is actively being developed. The use of QSAR for evaluation is progressing, such as the guideline (ICH-M7). Methods is roughly classified into knowledge-based judgments based on structures of concern for compounds, and statistically-based judgments that use computational scientific methods such as machine learning. ICH-M7 requires the use of both Ames QSAR assessments. Knowledge-based method makes judgments by establishing rules for chemical reaction mechanisms with DNA based on chemical reaction knowledge. Whereas, the compound structure input in silico is the structure before the metabolic reaction by the enzyme, and if the compound structure changes due to the metabolic reaction, the reactivity with DNA cannot be determined. Therefore, we constructed a function that simulates how the target compound is metabolized in the Ames test environment and affects the test results, lead to improvement in performance of knowledge-based prediction. Furthermore, we developed a prediction models trained on public and in-house Ames test data using GNN, a type of machine learning method, and combine it with the above knowledge-based model to make complementary judgments.
