Conformal MS2 Networks vs. Direct Toxicity Models: Pinpointing Endocrine-Disrupting LC–HRMS Features Without Structure ID

Environmental samples contain thousands of chemicals, resulting in a multitude of liquid chromatography-high resolution mass spectrometry (LC–HRMS) features, yet <10% are structurally elucidated, leaving most endocrine-disrupting compounds (EDCs) overlooked. Researchers at Stockholm University (Sweden) recently compared three MS²-driven frameworks that skip the necessity of structural elucidation —molecular networking, conformal predictors trained directly on spectra, and revised MS2Tox models—to prioritize features linked to endocrine-disruptive activity trained on seven Tox21 nuclear-receptor assays. Evaluated on wastewater influent/effluent, the approaches flag potential EDCs, and thus assist in processing non-target screening data.

LCGC International spoke to Yvonne Kreutzer, lead author of the paper that discusses the group’s research.¹

What chromatographic challenges do highly complex environmental water samples pose for LC-based non-targeted screening, and how can method development mitigate matrix effects and co-elution?

When analyzing complex water samples with liquid chromatography-high resolution mass spectrometry (LC-HRMS) in a non-target approach, often thousands of features are detected, but only a tiny fraction can be identified. Thus, most of the detected features remain unknown, and their toxic effects cannot be confirmed.To meet these challenges, we developed and compared post-acquisition data prioritization methods, enabling us to filter detected LC-HRMS features based on their probability of posing endocrine-disruptive activity.¹

How does chromatographic separation quality (for example, peak capacity, resolution, and retention time stability) influence the reliability of LC-HRMS feature detection and downstream MS²-based toxicity prediction?

Before predicting toxicity, we first pre-process acquired LC-HRMS data to extract chromatographic features. From here on, we continue with MS²-based toxicity predictions. The prediction accuracy heavily depends on the quality of recorded MS² spectra, and it has been shown before that low-quality MS² spectra can lead to incorrect fingerprint prediction with SIRIUS+CSI:FingerID.^2-4 However, on the other side SIRIUS+CSI:FingerID fingerprint prediction does not rely on retention times, making it resistant to retention time shifts. One important thing to keep in mind is that although we make a lot of progress being able to apply advanced processing methods to data from non-target screening, we cannot neglect the quality of the instrumental methods during sample acquisition.

Why is retention time considered a critical orthogonal parameter alongside exact mass and MS² spectra in non-targeted LC-HRMS workflows, and how is it used to reduce candidate structures?

Non-target analysis not only results in thousands of features, but when trying to identify features, each feature can also result in hundreds of potential candidate structures. In our group, we are also working on retention time predictions to reduce the number of resulting candidate structures. In practice, it will work like this: you first obtain your structural candidates from your workflow. For each of the candidate structures, a retention time is predicted based on its structure, and the predicted retention time is compared with the real one observed in your experiment. Given a user-specific error, you can then filter out candidates that do not match your experimental retention time, ultimately reducing the number of candidates.

What are the advantages and limitations of reversed-phase LC for separating endocrine-disrupting compounds with diverse polarity and functional groups in environmental samples?

Endocrine-disrupting compounds exhibit a wide range of polarities, and not all can be detected with reversed-phase LC. Still, reversed-phase LC allows us to separate and detect most of the endocrine-disrupting compounds we expect to find in aqueous samples sufficiently. More importantly, it can be easily coupled to electrospray ionization (ESI), which is most used within the non-target community. As we are interested in developing machine learning tools for data processing, one key aspect is data availability. Training a machine learning model does not only need high-quality data, but also a significant amount of data. This is especially critical for the data availability of MS² spectra for endocrine-disruptive chemicals in publicly available datasets. As LC-ESI-HRMS is the most common technique, most training data is available for this technique.

How can chromatographic fractionation be combined with bioassays to better link observed endocrine-disrupting activity to specific LC-HRMS features?

This is a good question, and co-workers in our group are collaborating in different projects to measure the in vitro bioassay activity of fractionated samples, while simultaneously applying in silico toxicity prediction models. Combining in silico and in vitro toxicities is a powerful combination for estimating how much of the experimentally observed in vitro toxicity can be explained with our model, ultimately enabling to estimate if toxic drivers of a samples have been identified. If you for example have measured the in vitro toxicity of a sample fraction, but this fraction still contains hundreds of features, in silico toxicity prediction enables to further prioritize and narrow down the features of interest.

In what ways does chromatographic co-elution affect MS² spectral quality, molecular networking, and the accuracy of ML models trained on MS² data?

Chromatographic co-elution, especially for chemicals with highly similar m/z values, has a negative impact on the quality of the MS² spectra, as the resulting fragmentation spectra are composites of several spectra. Such composite MS² spectra have a direct impact on similarity scores, such as the cosine score or MS2DeepScore. In molecular networking, they can cause incorrect linkages and subsequently lead to incorrect predictions. Similarly, low-quality spectra can affect the accuracy of predicted SIRIUS+FingerID fingerprints, and if fingerprints are predicted incorrectly, this error will propagate into subsequent toxicity modeling. Hence, co-elution may further affect the accuracy of prioritizing features related to toxic effects and is a major problem. Thus, different research projects also focus on how to evaluate MS² spectra quality.⁵

How might retention time alignment and reproducibility across LC-HRMS runs impact molecular networking and feature prioritization in large environmental datasets?

Aligning the retention time across LC/HRMS runs could have an impact on the spectral quality, as well as splitting data of one chemical into several features, and therefore downstream toxicity predictions on the obtained MS² spectra.

What considerations should be made when choosing LC versus gas chromatography (GC) for non-targeted screening of endocrine disruptors, given their physicochemical properties and required sample preparation?

Ultimately, it depends on which sample is being analyzed. In our case, we focus on water samples, as we all encounter water regularly, whether for drinking, cooking, or showering. Sample preparation for water samples for LC-HRMS screening is fortunately straightforward, as we can assume that most of our analytes are water-soluble. Therefore, a single filtration step may be sufficient, and the low sample preparation complexity minimizes analyte loss. In contrast, sample preparation for GC requires replacing the solvent with one that evaporates more easily, which involves several preparation steps during which some analytes could be lost.

How could non-target screening approaches be optimized to enhance the detection of low-abundance, but highly bioactive, endocrine disruptors in wastewater influent and effluent samples?

Low-abundance highly bioactive features are indeed a problem and may easily be overlooked in common non-target workflows. To address this, it is possible to predict the ionization efficiency of a detected feature and estimate its concentration, ultimately estimating the risk.⁶ However, one of the challenges that the community and general non-target workflows face is that if, during the chromatographic analysis no MS² spectra were triggered, the developed models cannot be applied.

Looking forward, how might advances in chromatographic techniques (such asmultidimensional LC or ion mobility–LC coupling) reduce uncertainty and error propagation in MS²-based toxicity prediction workflows?

Toxicity predictions, or also any kind of property prediction from MS² spectra, would benefit from using advanced chromatographic techniques. One of the main reasons is reduced co-elution and improved selectivity of very similar compounds, resulting in triggering more MS² spectra, which are also cleaner.⁷ In addition, more metadata is records, as for example a second retention time or a drift time, that can be used to narrow down candidate structures for feature identification.

References

1. Kreutzer, Y.; Rahu, I.; Norinder, U. et al. Molecular Networking, Conformal Predictions and Revised Fingerprint-Based Models for Discovering Endocrine Disruptors in Mixtures. Anal Bioanal Chem. 2026. DOI:10.1007/s00216-025-06303-2

2. Rahu, I.; Kull, M.; Kruve, A. Predicting the Activity of Unidentified Chemicals in Complementary Bioassays from the HRMS Data to Pinpoint Potential Endocrine Disruptors. J Chem Inf Model. 2024, 64, 3093–3104. DOI: 10.1021/acs.jcim.3c02050

3. Peets, P.; Wang, W-C.; MacLeod, M. et al. MS2Tox Machine Learning Tool for Predicting the Ecotoxicity of Unidentified Chemicals in Water by Nontarget LC-HRMS. Environ Sci Technol. 2022, 56, 15508–15517. DOI: 10.1021/acs.est.2c02536

4. Dührkop, K.; Fleischauer, M.; Ludwig, M. et al. SIRIUS 4: A Rapid Tool for Turning Tandem Mass Spectra into Metabolite Structure Information. Nat Methods 2019, 16, 299–302. DOI: 10.1038/s41592-019-0344-8

5. Codrean, S.; Kruit, B.; Meekel, N. et al. Predicting the Diagnostic Information of Tandem Mass Spectra of Environmentally Relevant Compounds Using Machine Learning. Anal Chem 2023, 95, 15810–15817. DOI: 10.1021/acs.analchem.3c03470

6. Sepman, H.; Malm, L.; Peets, P. et al. Bypassing the Identification: MS2Quant for Concentration Estimations of Chemicals Detected with Nontarget LC-HRMS from MS ² Data. Anal Chem. 2023, 95, 12329–12338. DOI: 10.1021/acs.analchem.3c01744

7. Hollender, J.; Schymanski, E. L.; Ahrens, L. et al. NORMAN Guidance on Suspect and Non-Target Screening in Environmental Monitoring. Environ. Sci. Eur. 2023, 35, 75. DOI: 10.1186/s12302-023-00779-4