Evaluation of Nontargeted\nMass Spectral Data Acquisition\nStrategies for Water Analysis and Toxicity-Based Feature Prioritization\nby MS2Tox

The machine-learning tool MS2Tox can prioritize hazardous\nnontargeted\nmolecular features in environmental waters, by predicting acute fish\nlethality of unknown molecules based on their MS<sup>2</sup> spectra,\nprior to structural annotation. It has yet to be investigated how\nthe extent of molecular coverage, MS<sup>2</sup> spectra quality,\nand toxicity prediction confidence depend on sample complexity and\nMS<sup>2</sup> data acquisition strategies. We compared two common\nnontargeted MS<sup>2</sup> acquisition strategies with liquid chromatography\nhigh-resolution mass spectrometry for structural annotation accuracy\nby SIRIUS+CSI:FingerID and MS2Tox toxicity prediction of 191 reference\nchemicals spiked to LC-MS water, groundwater, surface water, and wastewater.\nData-dependent acquisition (DDA) resulted in higher rates (19–62%)\nof correct structural annotations among reference chemicals in all\nmatrices except wastewaters, compared to data-independent acquisition\n(DIA, 19–50%). However, DIA resulted in higher MS<sup>2</sup> detection rates (59–84% DIA, 37–82% DDA), leading\nto higher true positive rates for spectral library matching, 40–73%\ncompared to 34–72%. DDA resulted in higher MS2Tox toxicity\nprediction accuracy than DIA, with root-mean-square errors of 0.62\nand 0.71 log-mM, respectively. Given the importance of MS<sup>2</sup> spectral quality, we introduce a “CombinedConfidence”\nscore to convey relative confidence in MS2Tox predictions and apply\nthis approach to prioritize potentially ecotoxic nontargeted features\nin environmental waters.