Accelerating Food Safety Analysis with an Open-Access LC–HRMS/MS Spectral Library

Ensuring food safety requires accurate detection of hazardous substances in complex food matrices and liquid chromatography–high-resolution tandem mass spectrometry (LC–HRMS/MS) provides an effective approach to achieve this. Unlike targeted analyses that detect known compounds, untargeted LC–HRMS comprehensively profiles all detectable analytes, but the vast amount of resulting data necessitates reliable tools for compound annotation and identification. Spectral library searching, matching experimental MS/MS spectra to reference spectra, is one of the most common annotation approaches. While several spectral libraries (including GNPS, MassBank, MoNA, and MSnLib) exist, none are specifically dedicated to food toxicants.

To bridge this gap, researchers at Wageningen Food Safety Research (WFSR), part of Wageningen University & Research (The Netherlands) developed the WFSR Food Safety Mass Spectral Library^,” a manually curated open-access LC–HRMS/MS spectral library containing 1001 food toxicants and 6993 spectra across multiple collision energies. The library encompasses diverse compound classes such as organic contaminants, natural toxins, pesticides, and veterinary drugs, among which also metabolites. By making this resource publicly available through platforms like GNPS and a dedicated webpage (Food Safety Mass Spectral Library - WUR), the researchers aim to improve the speed and accuracy of food toxicant identification, support structural analogue detection, and enable advanced applications such as machine learning–based structure prediction for food safety monitoring. LCGC International spoke to Ivan Aloisi, scientist at WFSR about this library, and the paper (1) which resulted from its development.

What advantages does LC–HRMS/MS offer for detecting chemical hazards in complex food matrices compared to other analytical methods?

LC–HRMS/MS is a powerful hyphenated technique playing a pivotal role in the analysis of food toxicants. LC is suitable for analytes having a wide range of polarities, non-volatile and thermally labile, such us most of the pesticides, per- and polyfluoroalkyl substances (PFAS), natural toxins, and veterinary drugs. HRMS provides exceptional resolving power, enabling accurate discrimination of analytes from matrix interferences in complex samples. This is particularly valuable in food safety applications, where food toxicants often occur at trace levels (in the low ppb or sub-ppb range), requiring highly sensitive and accurate analytical methods for reliable data generation and interpretation. Another important benefit of LC–HRMS/MS is that it enables the combination of suspect and untargeted screening with targeted analysis, allowing for a more comprehensive assessment of samples. Furthermore, this approach allows for retrospective data analysis, in which previously acquired datasets can be reprocessed as new toxicants are discovered, regulatory priorities shift, or to investigate the source and early occurrence of specific contaminant or group of contaminants. However, the application of complementary analytical techniques, such as gas chromatography for analytes like mineral oil saturated and aromatic hydrocarbons (MOSH and MOAH), and most of the persistent organic pollutants (POPs) such us dioxins, as well as inductively coupled plasma mass spectrometry (ICP-MS) for metals, is essential to comprehensively address the wide spectrum of food toxicants.

Can you explain the main differences between targeted and untargeted LC-HRMS approaches and when each is appropriate for food safety applications?

Targeted LC–HRMS screening relies on the injection of analytical reference standards at least once to gather retention time and diagnostic ion data for confident identification. Additionally, LC–HRMS can be employed for full quantitative analysis when reference standards are incorporated as quality controls (QCs) in each acquisition sequence. Sample preparation is usually optimized to enhance extraction efficiency, while instrumental settings are fine-tuned to achieve optimal performance. As a result, this approach offers high precision and sensitivity.

The concept of untargeted LC–HRMS analysis, as I interpret it, aims to maximize coverage of detectable compounds within the limits of sample preparation, instrumentation, and data processing. Achieving this requires a generic, non-selective sample preparation protocol, along with equally broad instrumental conditions and indefinite data processing possibilities. Untargeted analysis enables chemical characterization of a sample (or set of samples) by comparison between different groups, such as matrix blank and sample extracts, treatment and control, or treated samples across multiple time points. A key advantage lies in its ability to reveal previously unknown or unforeseen features (HRMS signals) that may escape detection in targeted or suspect analyses based on predefined analytes lists. This makes, in general, untargeted approaches powerful tools for detecting emerging and previously unknown hazards in complex matrices, including food. However, it remains highly complex and time-consuming, and no standardized strategy currently exists to prioritize detected signals in a way that simplifies and accelerates data processing.

Suspect screening LC–HRMS analysis can be conceptually positioned between targeted and untargeted strategies. In this approach, samples are screened for a list of analytes of interest based on prior knowledge, without the need for the injection of reference standards. Annotation is typically achieved by matching accurate mass, isotopic patterns, and more recently, by employing tandem mass spectral libraries. This approach often yields a list of tentatively identified analytes, which can subsequently be confirmed (or rejected) through the injection of reference standards. Suspect screening combines the focus of targeted analysis with the flexibility of untargeted approaches, enabling comprehensive and retrospective identification of known and emerging contaminants. Despite its power, suspect screening is constrained by its dependence on existing knowledge and lack of quantitative accuracy. It is, therefore, best used in combination with targeted and untargeted approaches in food safety applications.

Targeted analysis is ideal for regulatory monitoring and quantification of known food toxicants, whereas suspect and untargeted LC–HRMS are best suited for exploratory studies aimed at discovering emerging or unknown food toxicants in samples. Combining the three approaches provides the most informative data generation. LC–HRMS instruments allow the full integration of targeted, suspect, and untargeted workflows within a single analytical run, thereby maximizing information yield and efficiency. In my view, in the not-too-distant future, an increasing number of institutions will adopt hybrid analytical approaches designed to maximize the amount of information obtained from each data file.

This transition is already underway at Wageningen Food Safety Research, although it will still take some time and adjustments before being fully implemented. Moreover, the generation of large datafiles, it not only be useful for the abovementioned reasons, but it will contribute to the expansion and refinement of artificial intelligence capabilities. At our institute, we are investing significant effort in integrating analytical chemistry and data science, particularly with evidence obtained from LC–HRMS data.

How does spectral library searching contribute to compound annotation in untargeted LC–HRMS analysis, and what are the limitations of this approach?

Spectral library searching plays a central role in the tentative identification of analytes in both untargeted and suspect screening workflows. A spectral library is a collection of reference spectra that includes MS/MS fragmentation patterns, retention times (when available), together with other metadata of known compounds. It serves to match experimental with reference MS/MS spectra, allowing for the tentative identification of compounds.

Spectral libraries can be applied in targeted, suspect, and untargeted workflows to assist in compound identification. This strategy significantly enhances the confidence of annotation compared to approaches relying solely on accurate mass or isotopic patterns. The ability to isolate a precursor ion within a very narrow window (±0.8 m/z) and obtain from its fragmentation a characteristic spectral fingerprint increases the confidence of identification, particularly when dealing with highly complex matrices such as food.

Some key limitations include the possibility of false positives or inconclusive annotations caused by structural analogs or isomers producing similar spectra. Also, the fact that only a small portion of the chemical space is currently represented in available spectral libraries, and the lack of detailed metadata, particularly chemical class and retention time, makes it difficult to effectively filter and interpret the results. Spectral libraries can also be integrated into other untargeted workflows, such as molecular networking, in which compounds from the library may cluster with other features and thereby assist in the discovery of novel molecules. In the context of natural products, such as natural toxins, this approach is widely adopted and has proven valuable for elucidating potential new isomers and structural rearrangements. A recent scientific publication from our institute focused on building a dedicated spectral library of pyrrolizidine alkaloids (PAs) and using molecular networking to annotate novel PAs (2).

Why is the creation of a dedicated, open-access spectral library for food toxicants important, and what challenges does it address in current food safety workflows?

As of today, apart from the WFSR Food Safety Mass Spectral Library⁺,there is no other open-access spectral library that focuses exclusively on food toxicants and, consequently, on food safety. Moreover, the absence of standardized conditions, both in the chromatographic and mass spectrometric acquisition settings, necessitates extensive quality assessment of the information available in the online repositories. The absence of comprehensive metadata, for instance information on the chemical class to which the substances belong, makes it difficult to filter relevant food toxicants from large libraries such as the Global Natural Product Social Molecular Networking (GNPS) site.

The fact that 216 food toxicants included in the WFSR library are absent from other open-access repositories highlights its contribution to overcoming current challenges in food safety workflows. Another distinctive advantage of the Food Safety Mass Spectral Library⁺ is the inclusion of experimentally determined retention times for each food toxicants. This aspect is of particular significance, as it demonstrates that, provided the same chromatographic conditions are employed (see the experimental section in the publication [1]), all 1001 food toxicants included in the library can be successfully eluted and, consequently, analyzed and detected if present in any sample. The presence of metadata such as chemical classes, Chemical Abstract Services (CAS) numbers, Simplified Molecular Input Line Entry System (SMILES), and InChIKeys also facilitates pre- or post-processing filtering for specific classes of analytes and enables easy access to information regarding their chemical structure and spatial orientation. This represents a limitation of most existing open-access repositories, which require further improvement to achieve greater comprehensiveness.

The WFSR library includes spectra acquired at multiple collision energies and validated manually. How does this improve the reliability and usability of the data for compound identification?

For each food toxicant the HRMS/MS spectra were acquired at seven different collision energies thereby enabling their application (also retrospectively) to data files acquired at a specific collision energy. Noteworthy, one of the seven collision energy settings, known as stepped collision energy, applies three sequential collision energies that are merged into a single mixed spectrum. This approach accounts for differences in molecular bond stability and produces information-rich spectra with numerous fragment ions, thereby improving the robustness and interpretive power of spectral matching. Although the manual curation process was time-consuming and required careful attention, it proved essential to ensure that only well-characterized food toxicants were retained in the spectral library. During the curation stage, approximately 20% of the food toxicants were excluded due to noisy or low-quality mass spectra.

How might integrating the WFSR library with computational tools like GNPS or MS2Query enhance high-throughput screening and structural annotation of unknowns?

GNPS (3) and MS2Query (4) can compare thousands of experimental spectra with reference libraries, allowing researchers to quickly identify and screen analytes in complex samples. GNPS can also groups spectra into molecular networks based on their spectral similarity, enabling the identification of molecules that are structurally related to known toxicants present in WFSR Food Safety Mass Spectral Library⁺, even if exact matches are not present. This approach enhances detection capabilities beyond direct library matches. MS2Query uses machine learning to predict structural similarity between unknowns and library spectra, improving annotation of molecules that are not exact matches but share chemical substructures or fragmentation motifs, known as analogues. The WFSR Food Safety Mass Spectral Library⁺ was also used as part of the MS2Query training set for the improvement of a machine learning model for the detection of exact matches and structural analogues of known food toxicants. The re-training of the model was successfully finalized, but we still need to validate and evaluate it, hopefully soon.Integrating our spectral library with other computational tools, such as SIRIUS CSI:FingerID (5), can improve the prediction of molecular formulas, structures and chemical classes of analogues unknown from HRMS/MS spectra.

What role could machine learning play in improving compound identification from HRMS/MS data, and how might libraries like WFSR contribute to such advancements?

The use of machine learning is exponentially growing in analytical chemistry and in the food safety domain. The introduction of algorithms capable of learning from data to recognize patterns, classify molecules, and predict properties with minimal human intervention are transforming the way we approach food safety. However, machine learning models are highly dependent on the quality of the data used for their training. Therefore, ensuring high-quality data is essential to make these models truly effective and to enhance, rather than hinder, the overall analytical workflow. When combined with computational tools, the WFSR library can facilitates the identification of structural analogs and can also be applied in machine learning workflows to predict molecular structures or chemical classes. Thanks to the presence of retention time and structural identifier, it could also help in the prediction of retention time for other molecules that share similar chemical properties.

Given that 22.2% of compounds in the WFSR library are absent from other repositories, how could collaboration across institutions improve global coverage of food toxicants in spectral databases?

Collaborative efforts among institutions can significantly broaden the chemical coverage, enhance data quality, and promote open access to spectral information on food toxicants. Such cooperation helps unify fragmented datasets into a comprehensive global resource for the detection and surveillance of harmful substances in food. Because different parts of the world are exposed to distinct food types, contaminants, and environmental conditions, sharing data among institutions worldwide enables spectral databases to capture a broader range of toxicants, including region-specific or emerging contaminants that might otherwise go unnoticed. On a smaller scale, ongoing collaboration with other research groups within Wageningen University and Research aims to expand the library to encompass a broader range of chemical classes, such as polar food toxicants, persistent pollutants, and plant related toxicants. We are also open to discuss potential collaborations that could enhance the library’s coverage and contribute to improving food safety.

What future developments or expansions (for example, inclusion of negative ionization spectra) would you consider most valuable for improving the comprehensiveness of open-access spectral libraries in food safety research?

Regularly revising the spectral library on a yearly basis is crucial to ensure that it remains up to date and does not become obsolete. Expanding the library’s coverage to include additional ionization modes or chromatographic conditions can also be considered highly valuable improvements. In this regard, we are currently building a complementary spectral library containing food toxicants having superior ionization efficiency in negative mode, such as PFAS. This spectral library is scheduled for release by the end of the year. Future efforts will focus on expanding the scope towards highly polar food toxicants (such as glyphosate) through the development of a dedicated spectral library employing a dedicated chromatography separation.

Further updates will be made available through WFSR’s official website at the following link: Food Safety Mass Spectral Library - WUR.

References

1. Padilla-González, F.; Rizzo, S.; Dirks, C. et al. Creation of an Open-Access High-Resolution Tandem Mass Spectral Library of 1000 Food Toxicants. Anal. Chem. 2025, 97 (43), 23822-23830. DOI: 10.1021/acs.analchem.5c03020
2. Straub, L.V., Mulder, P.P.J., Zuilhof, H. et al. Development of a High-Resolution Tandem Mass Spectral Library for Pyrrolizidine Alkaloids (PASL). Sci. Data 2025. DOI: https://doi.org/10.1038/s41597-025-05940-7
3. Wang, M.; Carver, J. J.; Phelan, V. V; Sharing and Community Curation of Mass Spectrometry Data with Global Natural Products Social Molecular Networking. Nat. Biotechnol 2016. DOI: https://doi.org/10.1038/nbt.3597.
4. de Jonge, N. F.; Louwen, J. J. R.; Chekmeneva, E.; et al. MS2Query: Reliable and Scalable MS2 Mass Spectra-Based Analogue Search. Nat. Commun. 2023. DOI: https://doi.org/10.1038/s41467-023-37446-4.
5. Dührkop, K.; Shen, H.; Meusel, M.; et al. Searching Molecular Structure Databases with Tandem Mass Spectra using CSI:FingerID. Proceedings of the National Academy of Sciences (PNAS), 2015. DOI: https://doi.org/10.1073/pnas.1509788112