HTC-19 Preview: Predicting the Future — Non-Targeted Screening

Your talk in the Royal Society of Chemistry Sessions HTC-19¹ in Leuven, Belgium is entitled: NTS: Predicting The Future. For complex environmental matrices, how do you balance chromatographic separation efficiency with the risk of generating overwhelming numbers of high resolution mass spectrometry (HRMS) features for non-targeted screening (NTS) workflows in environemental analysis?

With increased chromatographic separation, the number of HRMS features with high-quality tandem mass spectrum tends to increase. This is true for both data-dependent (DDA) and data-independent acquisition (DIA) strategies.² In DDA, chromatographic peak resolution allows for more opportunities for quadruple isolation and fragmentation. In DIA, retention time-based deconvolution algorithms benefit from increased separation, and also the MS2 purity increases with fewer co-eluting precursor compounds. The increase in features with high-quality MS2 improves identification performance and also the prioritization used by many non-targeted screening (NTS) strategies. One of the main disadvantages is that the more features that are extracted may require longer compute times to process the information, especially for complex matrices. However, this burden is decreasing with the increasing availability of high-performance computing and the optimization of workflows, including artificial intelligence (AI)-assisted methods. The other disadvantage is that longer elution gradients required for better chromatographic separation (both one-dimensionsional [1D] and two-dimensional [2D]) are often the limiting rate for laboratories, where costs must be managed to meet project budgets and instrument time must be balanced with other users. As we push towards environmental monitoring studies that use HRMS around the world, these kinds of budget limitations will be disproportionately felt in the global south.

How effectively can retention time prediction or indexing be integrated into NTS pipelines to improve confidence in prioritising unknowns alongside tandem mass spectrometry (MS/MS)-based approaches?

Quantitative structure-retention relationship (QSRR) models have recently benefited from the open access publication of the METLIN reversed-phase liquid chromatography library, containing the experimentalretention time (RT) of over 80,000 analytical standards.³ Uncertainty in the prediction of RT and RT indices for unseen compounds within the models’ applicability domain (chemical space) typically falls between 5– 20% depending on which model is selected. This can also be negatively impacted by different chromatography conditions, like variable mobile phase buffers, column technology, temperature, and pressure. I have successfully implemented RT prediction filtering for suspect screening workflows to exclude features reducing the number of putative matches by up to 50%.⁴ However, the application of RT prediction for prioritization purposes in non-targeted screening (NTS) is still limited due to the requirement for confident structural annotation. Only a minor percentage of total features are typically annotated with high enough confidence to warrant RT prediction, thereby excluding the majority of features from the dataset. This is a very important area of research, and I think there may be some solutions to bypass the structural annotation, like we have seen with toxicity and bioconcentration.

To what extent does chromatographic co-elution compromise MS/MS spectral clarity, and how does this affect downstream processes like molecular networking or structural inference?

No two acquisition methods are quite alike, since researchers usually optimise their instruments and programs for specific aims and objectives. In addition to the advantages and disadvantages I previously mentioned with DIA and DDA, there are more nuanced acquisition settings that can influence the quality and quantity of features with high-quality MS2. For example, Sequential window acquisition of all theoretical mass spectrometry (SWATH-MS) (or more generally windowed DIA) can alleviate problems with the deconvolution of co-eluting peaks by splitting them into different scans as long as they have precursor mass-to-charge ration (m/z) differences that place them in different windows.⁵ But this needs to be balanced with the number of m/z windows and collision energy events, since this can increase the cycle time and reduce the number of scans per peak. Similarly, in DDA, users can tune the cycle time and number of collision events to change the number of precursor peaks that are able to be scanned with quadrupole isolation and fragmentation.⁶ The more features with high-quality MS2 annotation, the more molecular networks become able to identify features with similar structures and aids in identification.

Given the diversity of chromatographic methods, for example, reversed-phase liquid chromatography (RPLC) versus hydrophilic interaction chromatogrpahy (HILIC), how transferable are NTS prioritization strategies across different separation conditions?

All NTS prioritisation is completely transferable to each of the hyphenated HRMS separation techniques, including reversed-phase, HILIC, supercritical fluid chromatography (SFC) and ion mobility spectrometry (IMS) because they are statistical tools independent of the acquisition and feature extraction elements of the workflow. For example, you would need to perform feature extraction (peak picking, grouping and RT alignment) separately for samples acquired by RPLC and HILIC, or using positive and negative polarity modes. But the final feature lists from each extraction and polarity method could be merged to perform all NTS simultaneously. This may be looking for increasing or decreasing trends in feature abundance between samples from different locations or over time. With adequate adduct and chromatography annotation, combined batches could even be analysed by molecular networks (known as ion identity molecular networking). And many popular quantitative structure–activity relationship (QSAR)-based predictions are method agnostic.⁷ My own fishFingers model only requires structural fingerprints, derived from either SMILES or MS2, so would be able to rank and highlight chemicals with potential to bioaccumulate from all acquisition methods at once. This paper is in review at the moment but the link for the model can be found at https://github.com/drewszabo/fishFingers.Again, the increased cost in HRMS analysis seems to be the limiting factor, since each hyphenated method would require a separate injection.

While NTS aims to reduce bias in chemical prioritization, how much bias is still introduced at the chromatographic stage through compound retention (or lack thereof), especially for very polar or non-retained analytes?

Understanding the applicability domain (chemical space) covered by any workflow is one of the greatest challenges in hyphenated HRMS analysis. Sample collection and preparation are the first major obstacles for environmental analysis. There are issues of contamination from plastic and glassware, and the loss of analytes during filtration and extraction. Then, of course, chromatography will also be very selective in the range of substances that are (or are not) retained. And finally, the number of substances that can be ionized by any particular method limits the number even further. I would say that for every 1000 substances detected by a solid-phase extraction reversed-phase liquid chromatography electrospray ionizationhigh-resolution mass spectrometry [SPE–RPLC–ESI(+)–HRMS] method, there are at least another 1000 substances that were excluded by one of these steps.⁸ In reality, it’s likely that only a small fraction of substances are detected by a single method. To mitigate some of these issues, we have a range of quality control measures, like method blanks and duplicates that are fortified with analytical standards from a wide range of masses and polarities to evaluate the chemical space that the method is able to measure. And it’s critical that these biases are accounted for and reported in HRMS studies.

Have you explored incorporating chromatographic descriptors (e.g., retention behaviour, polarity indices) into models that predict bioaccumulation or toxicity from MS/MS data?

This is a really interesting question and something I have not really considered before. To be clear, the fishFingers model that I have trained and the MS2Tox model developed by my former group at Stockholm University (under Anneli Kruve) only use structural fingerprints to determine BCF and acute toxicity in fish, respectively.⁹ There are no inputs with explicit retention time order or indices. However, I will say for bioaccumulation models, there is a known relationship between hydrophobicity and bioconcentration factors. And in the interpretation of the structural fingerprints that contribute most to the fishFingers BCF model are strongly related to hydrophobicity, like halogenation, length of carbon chains, and presence of heteroatoms. In estimating the presence or absence of each of these structural fingerprints, programs like SIRIUS CSI:FingerID are used to calculate the probability from MS2 fragmentation trees.¹⁰ And to my knowledge, SIRIUS does not take retention time into account when deriving these structural fingerprints. This could be added to future versions and I would welcome this kind of change that would help improve the annotation performance of HRMS features.

References
1. NTS — Predicting the Future. Thursday 28 May 2026: 10;45 at HT-19. Website: https://htc-19.com
2. Szabo, D.; Falconer, T. M.; Fisher, C. M.; Heise, T.; Phillips, A. L.; Vas, G.; Williams, A. J.; Kruve, A. Online and Offline Prioritization of Chemicals of Interest in Suspect Screening and Non-targeted Screening with High-Resolution Mass Spectrometry. Anal. Chem. 2024, https://doi.org/10.1021/acs.analchem.3c05705
3. Domingo-Almenara, X.; Guijas, C.; Billings, E.; Montenegro-Burke, J. R.; Uritboonthai, W.; Aisporna, A. E.; Chen, E.; Benton, H. P.; Siuzdak, G. The METLIN small molecule dataset for machine learning-based retention time prediction. Nature Communications 2019, 10 (1), 5811. https://doi.org/10.1038/s41467-019-13680-7
4. Szabo, D.; Fischer, S.; Mathew, A. P.; Kruve, A. Prioritization, Identification, and Quantification of Emerging Contaminants in Recycled Textiles Using Non-Targeted and Suspect Screening Workflows by LC-ESI-HRMS. Anal. Chem. 2024, https://doi.org/10.1021/acs.analchem.4c02041
5. Xia, D.; Pan, G.; Liu, Y.; Liu, H.; Zhao, B.; Wu, J.; Tang, T.; Lu, G.; Wang, R. Unlocking the future potential of SWATH-MS: Advancing non-target screening workflow for the qualitative and quantitative analysis of emerging contaminants. Water Res. 2025, 277, 123323. https://doi.org/10.1016/j.watres.2025.123323
6. Davies, V.; Wandy, J.; Weidt, S.; van der Hooft, J. J. J.; Miller, A.; Daly, R.; Rogers, S. Rapid Development of Improved Data-Dependent Acquisition Strategies. Anal. Chem. 2021, 93 (14), 5676–5683. https://doi.org/10.1021/acs.analchem.0c03895
7. Matveieva, M.; Polishchuk, P. Benchmarks for interpretation of QSAR models. J. Cheminform. 2021, 13 (1), 41. https://doi.org/10.1186/s13321-021-00519-x
8. Zweigle, J.; Schlüsener, M.; Flottmann, J.; Bader, T.; Vidkjær, N. H.; Bollmann, U. E.; Christensen, J. H.; Tisler, S. Not One Method to Rule Them All: A Comparative Study of Chromatographic Platforms (RP-LC-, HILIC-, SFC-, and IC-HRMS) for Water Analysis. Anal. Chem. 2025, 97 (45), 25099–25110. https://doi.org/10.1021/acs.analchem.5c04114
9. Peets, P.; Wang, W.-C.; MacLeod, M.; Breitholtz, M.; Martin, J. W.; Kruve, A. MS2Tox Machine Learning Tool for Predicting the Ecotoxicity of Unidentified Chemicals in Water by Nontarget LC-HRMS. Environ. Sci. Technol. 2022, https://doi.org/10.1021/acs.est.2c02536
10. Dührkop, K.; Fleischauer, M.; Ludwig, M.; Aksenov, A. A.; Melnik, A. V.; Meusel, M.; Dorrestein, P. C.; Rousu, J.; Böcker, S. SIRIUS 4: a rapid tool for turning tandem mass spectra into metabolite structure information. Nat. Methods 2019, 16 (4), 299–302. https://doi.org/10.1038/s41592-019-0344-8