Abstract
The assessment of a compound’s potential DNA reactivity (mutagenicity) is a critical component of its safety evaluation. The biological activity of a compound is generally understood to be closely associated with its chemical structure. Because computer-aided quantitative structure-activity relationship (QSAR) models can provide high-throughput and rapid evaluation of compounds, the development of robust computational prediction models has substantial practical value. Herein, we developed the Decrypt Substance Toxicity (DSTOX) Mutagen Profiler software through a multi-strategy fusion approach. By integrating statistical rule-based models, expert knowledge rule-based models, analogue search, carcinogenicity data search, and International Council for Harmonisation (ICH) classification determination, the software generates mutagenicity assessment classification reports adhering to ICH M7 guidelines. Internal validation of the DSTOX statistical model indicated an accuracy of 92.0%. When combined with expert knowledge rule-based models and analogue cross-referencing, the DSTOX model exhibited robust mutagenicity prediction performance. External validation achieved an accuracy of 84.0% and a sensitivity (recall) of 82.8% for mutagenic compounds. The model’s performance metrics were generally comparable to those of current internationally recognized mainstream prediction tools.