After ASMS 2026, the Biggest Competitive Battleground in Untargeted Metabolomics Is Identification Confidence

The ASMS 2026 Signal

Walk through the poster hall at ASMS 2026 in Philadelphia this past June, and one trend was unmistakable: the conversation has shifted. For the past decade, metabolomics CROs competed on instrument specs — who had the highest resolution, the fastest scan speed, the most sensitive detector. Those conversations are fading. The new question researchers are asking at every booth is: "What percentage of the features you detect can you actually identify, and how confident should I be in those names?"

This shift reflects a maturing of the field. High-resolution mass spectrometers — Orbitrap Astral, timsTOF 2.0, the latest Q-TOFs — have reached a level of sensitivity where detecting 10,000 to 15,000 molecular features in a plasma sample is routine. But detecting a feature and knowing what it is are two entirely different achievements. A feature is just a pair of numbers: m/z and retention time. A metabolite has a structure, a pathway, a biological role. The gap between the two is what this article is about.

That gap is now the single most important factor separating a metabolomics CRO that delivers real biological insight from one that delivers a spreadsheet of anonymous peaks. After ASMS 2026, identification confidence is the competitive battleground — and in this article, we map the roadmap from feature list to credible biological insight.

Figure 1: The Identification Confidence Funnel

The 85% Problem: When Most of Your Data Has No Name

In a typical untargeted metabolomics experiment, somewhere between 10% and 20% of detected features get annotated at MSI Level 1 or Level 2 — meaning they have a name, a structure, and a reasonable level of confidence. The remaining 80% to 85% stay in what the field calls "metabolomic dark matter": features that are reliably detected, reproducibly quantified, and potentially biologically important, but chemically anonymous.

This is not a minor inconvenience. It has real consequences for every researcher who depends on untargeted metabolomics to answer biological questions:

For biomarker discovery, the top differentially abundant features between disease and control groups often turn out to be unknowns — classified only as "feature_3847" with an m/z value and a retention time. You know something changes, but you cannot map it to a pathway or build a biological narrative around it.

For drug mechanism studies, the metabolic shift you need to understand may be hiding among the unannotated features. If your untargeted dataset flags 200 significant features but only 30 have names, 85% of your potential mechanistic insight is inaccessible.

For multi-omics integration, the dark matter problem cascades. When you overlay transcriptomics and proteomics data onto a metabolomics layer where most signals are anonymous, the integrative power you hoped for collapses. You cannot connect a gene expression change to a metabolite that has no name.

The important thing to recognize is that this is not an instrument problem. The mass spectrometers are working beautifully. It is an identification strategy problem — and that is where the competitive landscape is being reshaped.

MSI Levels Are the Starting Line, Not the Finish Line

Every researcher working with metabolomics knows the Metabolomics Standards Initiative (MSI) framework, which defines four levels of identification confidence. The framework has been the foundation of metabolomics reporting since its introduction and was recently revisited in a comprehensive 2024 preprint by Metz, Fiehn, Wishart, and colleagues [1], who proposed augmenting the traditional qualitative MSI levels with an "identification probability" metric.

• MSI Level 1: Confirmed identification — matched to an authentic chemical standard under identical analytical conditions (same retention time, same MS/MS spectrum).
• MSI Level 2: Putatively annotated — high-quality match to a spectral library, but no in-house standard confirmation.
• MSI Level 3: Putatively characterized compound class — diagnostic fragment ions or neutral losses tell you it is a phosphatidylcholine or a glucuronide, but the exact structure remains uncertain.
• MSI Level 4: Unknown feature — detected and quantified, but structurally uncharacterized.

The MSI framework has served the field well for nearly two decades, but it has a fundamental limitation: it is a descriptive label, not a quantitative confidence measure. A Level 2 annotation could be backed by a spectral library match with a 95% cosine similarity score, or it could be a 62% match to the closest candidate in a small database. Both are reported as "Level 2." More importantly, the traditional MSI framework provides no estimate of the false discovery rate (FDR) — the proportion of your annotations that are likely wrong.

In proteomics, FDR control via target-decoy searching has been standard practice for over fifteen years. You report a list of protein identifications filtered to 1% FDR, and reviewers accept that as the benchmark. In metabolomics, until very recently, the equivalent was not possible. You annotated your features, assigned MSI levels, and hoped for the best.

In 2025 and 2026, this is finally changing — and the change has major implications for how researchers should evaluate metabolomics service providers.

FDR-Controlled Annotation: The DecoID2 and MassID Revolution

In February 2026, a team from Panome Bio and Washington University led by Gary Patti and Kevin Cho posted a preprint describing MassID, a cloud-based untargeted metabolomics pipeline with a module called DecoID2 that performs FDR-controlled metabolite identification [2].

The concept, adapted from proteomics, works like this: a database of decoy compounds (structures that do not exist in nature) is mixed into the annotation pipeline. The rate at which decoys are incorrectly identified as real metabolites provides an empirical estimate of the false discovery rate. By setting an FDR threshold — say, 5% — you can filter your annotation list to a known error rate, exactly as proteomics researchers have done for years.

The MassID paper reported striking results from human plasma: over 4,500 metabolites structurally identified, with more than 1,200 passing the 5% FDR threshold. Crucially, 884 of those high-confidence identifications came from what the traditional MSI framework would have classified as Level 2 or Level 3 compounds — annotations that most researchers would treat as provisional, but that turned out to be correct when subjected to FDR control. These compounds were always right; the field just lacked the statistical framework to prove it.

For a researcher choosing a CRO, the takeaway is direct and practical: a provider that can attach an identification probability to each annotation — and report results at a defined FDR threshold — is offering a fundamentally different product from one that delivers a flat list of MSI levels with no error-rate estimate. Between two CROs reporting similar numbers of "identified metabolites," the one with FDR control is the one whose list you can actually trust. This level of statistical rigor increasingly depends on sophisticated bioinformatics for metabolomics pipelines.

Figure 2: Traditional MSI Reporting vs. FDR-Controlled Reporting

Spectral Libraries: Bigger Is Not Always Better

Every untargeted metabolomics pipeline relies on spectral libraries — curated collections of MS/MS spectra from known compounds — to assign identities to detected features. The major public resources are impressive in scale: METLIN contains over 850,000 spectra, GNPS2 hosts a growing community-driven repository with ion identity molecular networking capabilities [5] and a repository-scale propagated suspect spectral library of over 87,000 annotated spectra [6], MoNA provides open-access spectra, and HMDB 5.0 maintains highly curated human metabolome entries [8]. Commercial libraries like mzCloud add another layer.

But counting entries tells you very little about how well a library will perform on your samples. Three factors matter more than raw size:

Multi-adduct coverage. Most spectral libraries are built around protonated ions [M+H]⁺. But in real biological samples analyzed by electrospray ionization, metabolites also form sodium adducts [M+Na]⁺, potassium adducts [M+K]⁺, ammonium adducts [M+NH₄]⁺, and others. If the library only contains [M+H]⁺ spectra and your metabolite of interest ionizes predominantly as [M+Na]⁺, the library match will be weak or absent — even though the compound is in the database. A quality library covers multiple adduct forms for each entry. The ion identity molecular networking approach introduced by Schmid and colleagues directly addresses this problem by connecting different ion species of the same molecule within molecular networks [5].

Experimental retention time. An MS/MS spectral match alone, even a good one, is not enough for MSI Level 1 confidence. You need retention time (RT) matching against an authentic standard run on the same chromatographic system. CROs that maintain their own in-house libraries with experimentally measured RT values — rather than relying solely on public databases — have a significant advantage in delivering Level 1 identifications. This is the strategy adopted by providers who explicitly prioritize "small but fully verified" over "large but uncurated."

Taxonomic relevance. A library of 200,000 plant natural products is less useful for a human plasma study than a focused library of 5,000 well-characterized mammalian metabolites. The COCONUT 2.0 database, released in 2025, addresses this by providing taxonomic source information for natural products, enabling organism-aware filtering during annotation [10]. If your CRO cannot explain how their library matches your study organism, that is a red flag.

A well-designed LC-MS/MS untargeted metabolomics workflow should integrate multiple spectral libraries and apply organism-relevant filtering — not just take the top hit from the largest available database.

CCS and Ion Mobility: The Fourth Dimension of Identification

If mass accuracy and MS/MS matching are the two traditional pillars of metabolite identification, collision cross section (CCS) from ion mobility spectrometry is rapidly becoming the third pillar — or, in the full 4D metabolomics workflow, the fourth dimension alongside retention time, accurate mass, and fragmentation spectra.

CCS measures how much a gas-phase ion is slowed by collisions with a buffer gas. It reflects the ion's shape — its three-dimensional conformation in the gas phase. Because CCS is a physical property of the ion, independent of the chromatographic system and the matrix, it provides a uniquely orthogonal piece of evidence for identification. Two isomeric metabolites that produce nearly identical MS/MS spectra can often be separated by their CCS values. The power of this approach has been demonstrated in lipidomics, where 4D trapped ion mobility enables high-throughput clinical profiling from minimal sample amounts [12].

A 2025 study from Leontyev and colleagues at Georgia Tech demonstrated the practical value of high-accuracy CCS for spatial metabolomics annotation. Using cyclic ion mobility with a CCS error below 0.4%, they showed that the CCS filtering window could be narrowed from the traditional ±3% down to ±1%, eliminating a much larger fraction of false candidate structures from SIRIUS/CSI:FingerID results [13]. In several cases, isobaric lipids were correctly annotated using CCS alone, even without MS/MS data. Complementary work from Huber and colleagues showed that trapped ion mobility CCS values improve annotation accuracy in exposomics screening applications, with their CCS library of 948 values helping resolve co-eluting matrix interferences across multiple biological matrices [11].

For a researcher evaluating untargeted metabolomics services, the CCS question is increasingly important. Does the CRO collect ion mobility data? What IMS platform do they use — drift tube (DTIMS, the gold standard for CCS accuracy), trapped ion mobility (TIMS), or cyclic IMS? Do they report CCS values alongside m/z and RT in their data deliverables? A CRO that integrates CCS into its identification workflow is operating at a different level of confidence than one that relies solely on mass accuracy and spectral matching.

Figure 3: 4D Metabolomics — How CCS Separates True Identifications from False Positives

The Multi-Evidence Stack: MS-Net, SIRIUS, and Why One Tool Is Not Enough

If there is one takeaway from the identification confidence research of 2025-2026, it is that no single tool gets it right often enough to rely on alone. Every computational annotation tool — whether it is spectral library matching, in silico fragmentation prediction, or molecular networking — has a characteristic error profile. The winning strategy is to stack them and look for consensus.

This is the principle behind MS-Net, a multi-similarity network annotation workflow described in a 2025 preprint by researchers at the Université de Toulouse. MS-Net integrates three layers of evidence: MS/MS spectral similarity (cosine scores from molecular networking), molecular structure similarity (Tanimoto scores comparing full-molecule and scaffold fingerprints), and taxonomic knowledge (filtering candidates through COCONUT 2.0 to retain only compounds consistent with the organism under study) [4]. The three layers are combined into a composite link score, and high-confidence annotations — those with strong spectral library matches or authentic standard confirmations — serve as seeds that propagate confidence through the network.

The headline result from MS-Net's proof-of-concept study on Cannabis sativa is worth pausing on: of 1,275 final confident annotations, 53% were "rescued" from lower positions in the initial in silico rankings. These were compounds that SIRIUS or MS-Finder had ranked at positions 2 through 50 — not wrong, just outscored by structurally similar but incorrect candidates. The multi-evidence approach corrected the ranking by incorporating information that no single tool had access to.

SIRIUS with CSI:FingerID remains the leading in silico fragmentation prediction platform, and its CANOPUS module provides compound class predictions that are valuable for MSI Level 3 assignments [9]. But used alone, its top-1 recommendation accuracy has well-known limitations. The optimal annotation workflow for a serious untargeted metabolomics study integrates feature detection (MS-DIAL 5 [7] or MZmine), followed by spectral library search (GNPS2 [5,6], MoNA, METLIN), then in silico prediction (SIRIUS/CSI:FingerID), multi-similarity network propagation (MS-Net [4]), taxonomic plausibility filtering (COCONUT 2.0 [10]), and FDR estimation (DecoID2 or equivalent [2]), ultimately yielding a confident annotation list.

A CRO's bioinformatics for metabolomics capability should be evaluated by how many of these layers they actually deploy — not by which single software package they license.

Figure 4: The Multi-Evidence Annotation Stack

MCheM and Chemical Labeling: Extracting Structural Information Beyond the Spectrum

For the most stubborn fraction of metabolomic dark matter — metabolites whose MS/MS spectra are simply not informative enough for confident annotation — a creative solution emerged in mid-2025 from an international collaboration led by Daniel Petras at the University of Tübingen.

MCheM (Multiplexed Chemical Metabolomics), published in Nature Communications in July 2025, uses three orthogonal post-column derivatization reactions performed online — meaning the chemical reactions happen between the LC column and the mass spectrometer, without manual sample handling [3]. The three reactions target different functional groups: L-cysteine reacts with electrophiles (Michael acceptors, epoxides, beta-lactones), AQC tags amines and phenols, and hydroxylamine captures aldehydes and ketones.

When a metabolite reacts with any of these reagents, its mass shifts by a predictable amount — and, critically, its MS/MS fragmentation pattern changes in ways that encode structural information about the functional group that reacted. Because the three derivatization reactions are chemically orthogonal, a single metabolite carrying multiple functional groups can produce up to three distinct derivatized products, each providing an independent line of structural evidence. This multiplexed readout is what gives MCheM its power: rather than relying on one fragmentation spectrum, the pipeline cross-references multiple spectra from the same parent compound, each highlighting a different reactive moiety.

The implementation is surprisingly practical. The derivatization reagents are infused post-column via a simple T-junction before the electrospray source, with reaction times on the order of seconds. No offline incubation, no fraction collection, no extra sample preparation. The enriched spectra are processed through a custom MZmine module called "Online Reactivity" and integrated with CSI:FingerID and GNPS2 for annotation. Across multiple test datasets spanning microbial extracts, human plasma, and environmental samples, MCheM improved annotation rates by 32% to 49%, with the method validated on 359 natural product standards and applied to the discovery of a novel glycosylated oxazolomycin from Streptomyces libani.

MCheM is a research-grade method, not yet a standard CRO offering. But its existence signals where the field is heading: the hardest-to-identify metabolites will not be solved by better databases alone. They will require chemical approaches that actively probe molecular structure — and the CROs investing in this kind of unknown metabolite identification capability today will define the competitive landscape in three to five years.

The Three-Tier Confirmation System: QC, Standards, and Targeted Verification

Identification confidence is not only a software problem. It is equally an experimental design problem, and the best computational pipeline in the world cannot rescue data from a poorly controlled analytical run. The CROs that consistently deliver high-confidence identifications operate what can be described as a three-tier confirmation system:

Tier 1 — QC-based feature reliability. Before asking whether a feature can be identified, ask whether it can be trusted. A pooled QC sample — prepared by mixing equal aliquots from all study samples — is injected every 6 to 10 runs throughout the batch. Features whose intensity varies by more than 20-30% (CV) across QC injections are flagged as analytically unreliable. Dilution QC (dQC) samples, an emerging best practice, go further: by analyzing serial dilutions of the pooled QC, the CRO can verify that each feature's intensity responds linearly to concentration. A feature that fails this linearity test is unlikely to represent a real metabolite. The mQACC consortium, which published consensus best-practice recommendations for QA/QC in LC-MS-based untargeted metabolomics, considers pooled QC monitoring the minimum standard for any publication-grade study [14].

Tier 2 — Authentic standard verification. For the subset of features that emerge as candidates for biological interpretation — the top differentially abundant features in a biomarker study, or the metabolites that define a pathway-level shift — verification against an authentic chemical standard is the only route to MSI Level 1 confidence. This means running the pure standard on the same LC-MS/MS system under identical conditions and confirming that both retention time and fragmentation spectrum match. The size of a CRO's in-house standard library directly limits how many Level 1 identifications they can deliver. As Köfeler and colleagues established in their lipidomics annotation quality framework, correct annotation requires more than matching spectra — it demands attention to adduct forms, retention time modeling, and head-group-specific fragments that can distinguish otherwise similar structures [15].

Tier 3 — Targeted assay confirmation in an independent cohort. For the highest-stakes findings — biomarker candidates destined for publication or downstream validation — the gold standard is to develop a targeted MRM (multiple reaction monitoring) or PRM (parallel reaction monitoring) assay for the specific metabolite, then measure it in an independent set of samples. This is the metabolomics equivalent of what the ICH M10 guideline describes for bioanalytical method validation, and it transforms a "putative" finding into a confirmed result. A CRO that offers both untargeted discovery and targeted metabolomics under one roof can take a biomarker from discovery through verification without the data loss and delays that come from switching providers between phases.

The majority of CROs operate at Tier 1 only. Some reach Tier 2. Very few can execute all three tiers in an integrated metabolomics service workflow — and that capability gap is exactly what "identification confidence" as a competitive differentiator looks like in practice.

Figure 5: The Three-Tier Confirmation Pyramid

From Feature Table to Biological Insight: A Six-Step Roadmap

Pulling together everything discussed so far, the journey from a raw untargeted metabolomics dataset to a set of biologically meaningful, defensibly identified results follows six sequential steps. Each step builds on the one before it, and skipping any step introduces a vulnerability into the final conclusions:

Step 1 — Acquisition. Start with 4D metabolomics: retention time, accurate mass, MS/MS fragmentation spectra, and — wherever the platform permits — collision cross section. The more orthogonal dimensions of data per feature, the more constraints are available for downstream identification.

Step 2 — Annotation. Run the multi-evidence stack: spectral library search (METLIN, GNPS2, MoNA), in silico prediction (SIRIUS/CSI:FingerID), and network-based propagation (MS-Net). Filter by taxonomic relevance (COCONUT 2.0). Apply FDR estimation (DecoID2 or equivalent) to produce a probability-ranked annotation list, not just MSI level assignments.

Step 3 — Filtering. Remove features that fail QC criteria: CV too high in pooled QCs, poor linearity in dQC samples, high signal in method blanks. Quantify and report batch effects, and apply batch correction (QC-RLSC, SERRF, or ComBat) with transparent documentation of the method and parameters used.

Step 4 — Verification. For the subset of identified features that will drive the biological narrative, confirm identity against authentic standards. Report the standards used, the RT match tolerance, and the MS/MS spectral similarity score for each verified metabolite. If your key candidate is a novel compound without a commercial standard, dedicated unknown metabolite identification using high-resolution MSⁿ or preparative isolation becomes necessary.

Step 5 — Contextualization. Map verified metabolites to biological pathways (KEGG, Reactome, MetaCyc), perform enrichment analysis, and connect metabolite changes to the study's overarching biological question. This is where bioinformatics for metabolomics capability becomes critical — the most interesting biology often hides among metabolites not yet in standard pathway databases.

Step 6 — Validation. For biomarker studies and other high-stakes applications, confirm the top candidates by targeted metabolomics using MRM or PRM in an independent sample cohort. This closes the loop from discovery to verification and gives reviewers — and you — the confidence that the findings are real.

A researcher who understands these six steps is equipped to evaluate CRO capabilities at a level of detail that goes far beyond comparing instrument lists and price quotes. The question is not "what mass spectrometer do you use?" but "how many of these six steps does your pipeline actually execute, and can you show me the data for each one?"

Practical Takeaways for Researchers

If you are planning an untargeted metabolomics study in 2026, here are the questions that matter:

"What is your typical identification rate at MSI Level 1 versus Level 2?" A CRO that cannot answer this question with real numbers from projects similar to yours has not been tracking it — and you cannot improve what you do not measure.

"Do you provide FDR estimates for your annotations?" In 2026, the answer should be yes. If it is not, ask what their timeline is for implementing FDR-controlled annotation.

"What tools are in your annotation pipeline?" A single-tool pipeline is a red flag. Look for evidence of multi-evidence integration: library matching plus in silico prediction plus network-based propagation.

"Do you collect CCS data, and do you use it to filter candidate structures?" If ion mobility is available on their platform, CCS should be part of the identification workflow, not an unused data channel.

"Can you take a candidate biomarker from untargeted discovery through targeted verification without switching providers?" An integrated untargeted-to-targeted workflow — like a comprehensive metabolomics service that covers the full pipeline — is one of the strongest signals of a mature metabolomics operation.

The identification confidence gap is not going away — but it is becoming solvable. The tools exist. The statistical frameworks exist. The chemical approaches exist. What separates the CROs on the right side of this gap from those on the wrong side is whether they have invested in deploying them as an integrated pipeline rather than a checklist of individual capabilities. If your next study demands answers, not anonymous features, the time to ask about identification confidence is before the samples ship.

FAQ

Can I publish MSI Level 2 identifications in a high-impact journal?

Yes — Level 2 annotations are the workhorse of most published metabolomics studies. The key is transparency: report exactly which library was matched, the spectral similarity score, and whether any orthogonal evidence (CCS, RT plausibility) supports the assignment. What reviewers increasingly push back on is Level 2 annotations presented as if they were definitive.

How many Level 1 identifications do I need for a convincing study?

There is no universal number. A study that identifies 50 metabolites at Level 1 and maps them to a coherent pathway story is usually more compelling than one reporting 500 Level 1 identifications with no biological narrative. Focus on what the identified metabolites tell you, not the count.

Does FDR-controlled annotation cost more than traditional MSI-based reporting?

As of 2026, the tools for FDR-controlled annotation (like DecoID2) are still emerging from academic labs, and not all untargeted metabolomics providers offer them. Where available, it may involve additional bioinformatics time. The value, however, is in reducing the risk that your key finding is a false annotation — a risk that can be far more expensive than any analysis fee if it leads to a retraction or a failed follow-up study.

Is ion mobility useful for all types of metabolites?

Ion mobility provides the most value for metabolites in the mid-polarity range, including lipids, where isomeric forms are common and CCS differences are informative. For very small, rigid molecules, the CCS differences may be too subtle to add much discriminating power. This is especially relevant in lipidomics applications, where CCS has become a key parameter for resolving structurally similar lipid species. A good CRO will advise on whether ion mobility adds value for your specific compound classes of interest.

What if my key biomarker is still unknown after all of this?

This is common and not a failure. Unknown metabolites identification is a specialized workflow that uses additional techniques — high-resolution MSⁿ, chemical derivatization, preparative-scale isolation followed by NMR — to characterize novel compounds. If a differentially abundant unknown feature is central to your biological story, invest in dedicated unknown identification rather than accepting "feature_3847" in your paper.

What does a publication-ready untargeted metabolomics dataset include?

At minimum: raw data files (vendor or open format like mzML), a feature quantification matrix with CV values from pooled QCs, the annotation list with MSI levels and supporting evidence scores for each metabolite, and a description of the data processing parameters. Increasingly, deposition of raw MS data to MetaboLights or MassIVE is expected by reviewers. If ion mobility data were acquired, CCS values should be included. The goal is that another researcher could reproduce your analysis from the deposited files.

References

1. Metz TO, Chang CH, Gautam V, et al. Introducing "identification probability" for automated and transferable assessment of metabolite identification confidence in metabolomics and related studies. bioRxiv [Preprint]. 2024. DOI: 10.1101/2024.07.30.605945
2. Stancliffe E, Gandhi T, Guzior DV, et al. MassID provides near complete annotation of metabolomics data with identification probabilities. bioRxiv [Preprint]. 2026. DOI: 10.64898/2026.02.11.704864
3. Vitale GA, Xia SN, Dührkop K, et al. Enhancing tandem mass spectrometry-based metabolite annotation with online chemical labeling. Nature Communications. 2025;16:6911. DOI: 10.1038/s41467-025-61240-z
4. MS-Net: Multi-Similarity based network annotation for untargeted metabolomics. Research Square [Preprint]. 2025. DOI: 10.21203/rs.3.rs-8174529/v1
5. Schmid R, Petras D, Nothias LF, et al. Ion identity molecular networking for mass spectrometry-based metabolomics in the GNPS environment. Nature Communications. 2021;12:3832. DOI: 10.1038/s41467-021-23953-9
6. Bittremieux W, Avalon NE, Thomas SP, et al. Open access repository-scale propagated nearest neighbor suspect spectral library for untargeted metabolomics. Nature Communications. 2023;14:8488. DOI: 10.1038/s41467-023-44035-y
7. Tsugawa H, et al. MS-DIAL 5 multimodal mass spectrometry data mining unveils lipidome complexities. Nature Communications. 2024;15:7934. DOI: 10.1038/s41467-024-54137-w
8. Wishart DS, Guo A, Oler E, et al. HMDB 5.0: the Human Metabolome Database for 2022. Nucleic Acids Research. 2022;50(D1):D622-D631. DOI: 10.1093/nar/gkab1062
9. Blaženović I, Kind T, Ji J, Fiehn O. Software tools and approaches for compound identification of LC-MS/MS data in metabolomics. Metabolites. 2018;8(2):31. DOI: 10.3390/metabo8020031
10. Sorokina M, Merseburger P, Rajan K, et al. COCONUT 2.0: an updated and comprehensive database of natural products. Nucleic Acids Research. 2025. DOI: 10.1093/nar/gkae1063
11. Huber C, Ulrich EM, Krauss M. Trapped ion mobility improves annotation accuracy in LC-HRMS screening applications for exposomics. Analytical Chemistry. 2025. DOI: 10.1021/acs.analchem.5c00785
12. Lerner R, Baker D, Schwitter C, et al. Four-dimensional trapped ion mobility spectrometry lipidomics for high throughput clinical profiling of human blood samples. Nature Communications. 2023;14:937. DOI: 10.1038/s41467-023-36520-1
13. Leontyev D, et al. A spatial metabolomics annotation workflow leveraging cyclic ion mobility and machine learning-predicted collision cross sections. Journal of the American Society for Mass Spectrometry. 2025;36(6):1386-1394. DOI: 10.1021/jasms.5c00090
14. mQACC Consortium. Metabolomics 2023 workshop report: moving toward consensus on best QA/QC practices in LC-MS-based untargeted metabolomics. Metabolomics. 2024;20:35. DOI: 10.1007/s11306-024-02135-w
15. Köfeler HC, et al. Quality control requirements for the correct annotation of lipidomics data. Nature Communications. 2021;12:4771. DOI: 10.1038/s41467-021-24984-y