Read-across prediction of nongenotoxic thyroid carcinogenicity of pesticides in rats using mechanism-based in vitro assay data.

Developing alternatives to rat carcinogenicity studies remains a major challenge. In our previous study, we found that while structural similarity provided a basis for grouping carcinogens, integrating mechanistically relevant in vitro cytotoxicity data substantially increased the accuracy of read-across predictions for tumors in the nasal cavity, stomach, and bladder/urethra. In this study, using the same dataset of 126 pesticides, we investigated whether a toxicity mechanism-based read-across approach could be applied to predict thyroid carcinogenicity. We evaluated nuclear receptor activation, which is relevant to thyroid carcinogenic mechanisms, using in vitro assays. Read-across predictions were performed by first selecting neighboring substances based on Euclidean distances calculated from molecular descriptors, and then further refining them based on concordance in in vitro assay results. Our results demonstrated that selecting neighboring substances based on both molecular descriptors and in vitro assay data improved prediction accuracy (balanced accuracy: 0.722) compared to selecting substances based on molecular descriptors alone (balanced accuracy: 0.566). This suggests that combining molecular descriptors with mechanistically relevant in vitro assay data is effective for predicting thyroid tumors in rats, further supporting our integrated read-across approach for carcinogenicity prediction.