Transcending Structural Dependencies: A Tunable Mass Spectrometry-Driven Machine Learning Framework for Genotoxicity Prediction.

The escalating global challenge of genotoxic compounds (GCs) in environmental, pharmaceutical, and food contexts necessitates analytical approaches that combine high efficiency with independence from prior structural information. Here, we present GenoToxMass, a pioneering mass spectrometry-based predictive model that eliminates dependence on prior chemical structure. Constructed from a rigorously curated data set of 16,806 mass spectra and utilizing multimodal features, GenoToxMass achieves an area under the curve (AUC) of 0.95 with robust generalizability in external validation (AUC = 0.89). Real-world application demonstrates >85% concordance with Ames test outcomes and literature, correctly classifying 42 chemical products and 16 drug impurities in complex matrices. An openly accessible web platform enables practical deployment, while a Standard Operating Procedure (SOP) is provided for regulatory agencies. Interpretable spectral features and structural alerts enhance transparency beyond conventional "black-box" predictions, while model performance can be context-adapted through tunable decision thresholds. GenoToxMass establishes a versatile, scalable paradigm for the nontargeted screening of GCs across environmental, industrial, and regulatory landscapes, offering a new pathway that diverges from traditional animal-dependent toxicological methods.