Prioritizing Unknowns in Nontarget Screening of Food Samples

LCGC International spoke to Anneli Kruve of Stockholm University about nontarget screening of food samples by high-resolution mass spectrometry (HRMS), the subject of her RAFA 2026 talk, "Data-Driven Strategies to Reduce Complexity in Non-Target HRMS Analysis of Food Samples." Food matrices produce thousands of chemical features, most of them unidentified, so food safety laboratories risk spending interpretation time on features of low concern. Kruve details how predicted collision cross-section (CCS) values helped to remove 16% of incorrect candidate structures while retaining 90% of correct ones, and how retention time filtering can remove more than 40% of incorrect candidates. She advises laboratories to test bioactivity and concentration prediction models on known chemicals in their own sample types before relying on them for prioritization.

LCGC International: Your talk at RAFA 2026 highlights that food matrices contain thousands of features, but the vast majority remain unidentified. From a practical standpoint, what does that bottleneck actually cost a food safety laboratory—what decisions can't be made because of it?

Anneli Kruve: In laboratories practicing nontarget screening, it takes a lot of time to elucidate and annotate detected chemical features, particularly if none can be prioritized based on the metadata or analytical information. This means that valuable data interpretation time may be spent on chemical features that are not of the highest concern and, therefore, less effort remains for those that are potentially hazardous.

How reliable is predicting toxicity—like acute aquatic toxicity or endocrine activity—directly from an MS² spectrum, without knowing the compound's structure?

All machine learning/artificial intelligence (ML/AI) predictions come with uncertainty, acknowledged or not. Predictions from spectra are still less accurate than the predictions made from structures, which have larger uncertainty than experimental values. Nevertheless, experimentally determined toxicity end points also have a substantial uncertainty arising from biological variability, which—depending on the end point—may reach an order of magnitude. These uncertainties definitely depend on the end point, but also on how similar the detected compounds are to the ones used for model training. It is therefore not easy to give a certain number. The most important aspect is acknowledging the uncertainty and weighing this against the alternatives, rather than the number itself. For cases where structures are annotated confidently, it is definitely worth investigating predictions from structure or even checking availability of experimental activity data. However, if the structure elucidation is ambiguous, the predictions from spectra may be the only feasible option.

What did your recent comparison of conformal MS² networks, molecular networking, and direct toxicity models teach you about when to trust which method?

In our experience, the activity predictions made from molecular fingerprints predicted from MS² data show the highest accuracy.¹ This indicates that general structural similarity (captured in molecular networking and conformal predictions from spectral similarity) is somewhat less informative for activity predictions than molecular fingerprints, where specific fingerprint features may be associated with end point–specific molecular moieties.

How accurate is a concentration estimate based on predicted ionization efficiency, without a reference standard?

Performance depends on the compound class and the instrumental method (both liquid chromatography [LC] conditions and mass spectrometry [MS]). For water analyses, average fold errors for predictions from structure range from twofold to an order of magnitude.² Heeren et al also recently reported an average fold error of 2.5 (maximum 4.7) in bee pollen.³ 

How much do retention time and collision cross-section (CCS) actually reduce false positives when narrowing candidate structures in food matrices?

Both retention time and CCS predictions are increasingly used to remove/label unlikely candidate structures suggested in nontarget screening based on MS data. The exact performance strongly depends on the method used for candidate structure retrieval, the number of candidate structures considered, and the chemical space.⁴ Using our generic CCS prediction evaluations, we removed 16% of incorrect structures while retaining 90% of correct ones.⁵ Retention time filtering can remove more than 40%.

How mature are de novo structure generation tools right now, and what should someone watch out for when one claims a confident structure?

These tools are developing fast, and the next few years will show what is possible in the field. It is unlikely that these tools will replace spectral library matching or analytical standards for gaining higher confidence in the structural annotation; however, they will most likely expand our knowledge of the sample composition as more high-resolution mass spectrometry (HRMS) features will be annotated with likely structures.

What's a realistic first step for a food safety laboratory that wants to start using bioactivity- and concentration-based prioritization today?

For a laboratory already using nontarget screening, it would be to evaluate the performance of bioactivity and concentration prediction models on known chemicals and samples. It is important to know that the performance is acceptable for the sample types and groups of chemicals most relevant for the scope of the laboratory.

References

1. Kreutzer, Y.; Rahu, I.; Norinder, U.; Kruve, A. Molecular networking, conformal predictions and revised fingerprint-based models for discovering endocrine disruptors in mixtures. Anal. Bioanal. Chem. 2026, 418 (5), 1445–1457. DOI: 10.1007/s00216-025-06303-2
2. Malm, L.; Liigand, J.; Aalizadeh, R.; Alygizakis, N.; Ng, K.; Kjær, E. E. F.; Nanusha, M. Y.; Hansen, M.; Plassmann, M.; Bieber, S.; et al. Quantification approaches in non-target LC/ESI/HRMS analysis: An interlaboratory comparison. Anal. Chem. 2024, 96 (41), 16215–16226. DOI: 10.1021/acs.analchem.4c02902
3. Heeren, S. M.; Kramer, N. I.; Righetti, L. Integrating suspect and nontargeted screening for food safety testing: A case study of bee pollen supplements. Anal. Chem. 2026, 98 (29), 21732–21741. DOI: 10.1021/acs.analchem.6c02768
4. Hupatz, H.; Rahu, I.; Wang, W.-C.; Peets, P.; Palm, E. H.; Kruve, A. Critical review on in silico methods for structural annotation of chemicals detected with LC/HRMS non-targeted screening. Anal. Bioanal. Chem. 2025, 417 (3), 473–493. DOI: 10.1007/s00216-024-05471-x
5. Ferrando Plo, L.; Kruve, A. Candidate structure prioritization in non-target screening with predicted collision cross-sections and uncertainty quantification. J. Chem. Inf. Model. 2026. DOI: 10.1021/acs.jcim.6c02492