Untargeted metabolomics routinely detects thousands of features in a single LC-MS run, far more than manual review can interpret. AI untargeted metabolomics LC-MS workflows are now the only practical route to annotating that complexity at scale. The annotation rate for genuinely novel metabolites, however, remains low, and this article addresses that gap directly.
Key Takeaways
- AI tools such as molecular networking platforms and in silico fragmentation engines have become the standard route to metabolite annotation in untargeted LC-MS metabolomics.
- Even with AI-assisted annotation, most features detected in a typical untargeted experiment remain unidentified at the structural level.
- Deep learning peak curation reduces false-positive peaks from automated feature detection without discarding true chromatographic signal.
- Statistical and pathway analysis platforms convert annotated metabolite lists into candidate biomarkers and biological context, not confirmed results.
- The dark metabolome debate shows that unannotated features can reflect both genuinely novel chemistry and instrument artifacts, and the two are not easy to tell apart.
AI Untargeted Metabolomics LC-MS Analysis Starts With a Complexity Problem
A single untargeted LC-MS run can generate tens of thousands of mass spectral features from one biological sample. Most of those features correspond to adducts, isotopes, or in-source fragments rather than distinct metabolites, and separating real chemistry from analytical noise is the first problem any workflow has to solve before biology enters the picture at all. Manual review does not scale to this volume, which is one reason computational and machine learning methods have moved from optional to default across the field. Even well-resourced labs running dozens of samples per week cannot manually inspect every feature without a queue of unreviewed data building up behind them.
Working in analytical science?
Register for a FREE Separation Science account to subscribe to the Separation Science Newsletter.
The annotation gap is the more persistent problem, and it has not closed nearly as fast as feature detection has improved. On average, only about 10% of the molecules detected in a typical untargeted metabolomics experiment can be structurally annotated, a figure highlighted in a computational metabolite annotation review. That low annotation rate limits biochemical interpretation and makes comparing results across different metabolomics studies genuinely difficult, since two labs studying the same biology may annotate entirely different subsets of the same underlying chemical space.
AI does not close this gap by working harder at the same manual matching process analysts already use. It changes the underlying approach, applying pattern recognition across large spectral libraries and trained embedding models at a scale and consistency that no individual chemist could sustain across thousands of samples, a shift that mirrors the broader move toward AI in analytical science across chromatographic and spectroscopic techniques generally. The result is not complete annotation, but it is a meaningfully larger and more consistent annotated fraction than manual workflows alone can deliver.
Automated Feature Detection and Peak Alignment Reduce False Positives
Before any compound can be annotated, raw chromatographic and spectral data has to be converted into a reliable table of features. Established open-source peak-picking tools build this table by defining regions of interest, detecting chromatographic peaks within them, and aligning matched peaks across samples so that the same metabolite is tracked consistently from run to run. These algorithms have a well-documented tendency to overpick, however, generating large numbers of false-positive peaks alongside genuine signal, and one comparison found that peak lists from two widely used peak-picking tools run on the same dataset overlapped on as little as 36% of the peaks each tool reported.
Deep learning models trained specifically to distinguish true chromatographic peaks from noise have been shown to remove the large majority of these false peaks without discarding true-positive signal, an approach demonstrated by deep learning peak curation tools built on convolutional neural networks. Rather than relying on a fixed set of manually tuned intensity or shape thresholds, these models classify each candidate peak using the full profile of its chromatographic curve, which lets them generalize across instrument types and sample matrices more reliably than rule-based filters.
The practical benefit shows up downstream. A cleaner feature table entering annotation carries fewer artifacts that would otherwise be mistaken for real metabolites during spectral matching or statistical analysis, and it reduces the manual curation burden that has traditionally followed automated peak picking as a matter of course. Peak curation does not replace careful parameter optimization at the acquisition and processing stage, but it substantially narrows how much of that curation has to happen by eye.
AI-Assisted Metabolite Annotation Spans Databases and in Silico Fragmentation
Once a curated feature table exists, annotation typically proceeds through spectral library matching first, comparing each sample's fragmentation spectrum against reference libraries built from authentic standards or previously characterized compounds. When no library match exists, which is common given that reference libraries cover only a small fraction of known chemical space, in silico fragmentation methods step in to predict fragmentation patterns for candidate structures pulled from large chemical compound databases and rank those candidates by how well their predicted spectrum agrees with the one actually observed.
Molecular networking adds a further layer of context by grouping mass spectra with high structural similarity into clusters, so that the confirmed identity of one compound can help annotate its structurally related neighbors within the same cluster even without a direct library hit. Machine learning embeddings trained on large spectral datasets have improved this matching considerably, correlating more closely with true structural similarity than traditional cosine-based similarity scores, which historically struggled with spectra that were structurally close but not identical, a limitation covered in more general terms in machine learning in mass spectrometry data analysis.
A practical annotation workflow for untargeted LC-MS data generally follows a consistent sequence regardless of which specific tools a lab uses:
- Filter the aligned feature table to remove known artifacts, including isotopologues, adducts, and in-source fragments, before annotation begins.
- Run spectral library matching against curated reference databases as the first annotation pass, since it carries the highest confidence when it succeeds.
- Apply in silico fragmentation tools to features with sufficient MS/MS data that did not return a library match.
- Use molecular networking to propagate identities across structurally related spectral clusters and flag candidate analogues for manual review.
- Assign an explicit confidence level to every annotation before it enters statistical or biological interpretation.
Statistical Analysis Turns Annotated Features Into Candidate Biomarkers
Annotated feature tables feed into univariate and multivariate statistical analysis to identify which metabolites differ meaningfully between comparison groups. Multivariate methods, particularly partial least squares discriminant analysis and random forest classification, remain the workhorses for this task because they capture the correlated, high-dimensional structure of metabolomics data more effectively than univariate testing alone, a point developed at length in a statistical biomarker discovery methods review that compares the strengths and failure modes of each approach.
Once statistically significant metabolites are identified, pathway mapping tools translate that shortlist into biological context by testing for enrichment against curated metabolic pathway databases rather than leaving analysts to interpret a flat list of chemical names. Web-based platforms such as the MetaboAnalyst data analysis platform combine these statistical and pathway modules with receiver operating characteristic analysis for biomarker evaluation, all inside a single reproducible workflow that does not require a programming background to run.
The output of this stage is a shortlist of candidate biomarkers and affected pathways, not a validated diagnostic result, and treating it as anything more is a common and costly misstep. Every statistical hit still requires orthogonal confirmation, ideally against an authentic reference standard measured under matched conditions, before it can support a biological or clinical claim of any weight.
The metabolomics field has largely converged on the Metabolomics Standards Initiative (MSI) framework for describing how confident an annotation actually is, and every AI-assisted call should be labeled against it before it moves further downstream:
| Confidence level | What it requires | Typical role of AI |
|---|---|---|
| Confirmed identity | Matches an authentic reference standard measured under identical analytical conditions | Confirms candidates that AI tools flag rather than generating the identification itself |
| Probable structure | Matches a spectral library or in silico prediction without a reference standard on hand | Where spectral matching and fragmentation prediction do most of the annotation work |
| Putative class | Chemical class assigned from spectral similarity alone, without a specific structure | Where AI classification tools assign compound class from fragmentation patterns |
| Unknown | No structural information beyond mass and retention time | The remaining share of features, often referred to as the dark metabolome |
The Dark Metabolome Remains Beyond Reliable AI Annotation
Even with library matching, in silico fragmentation, and molecular networking applied together, a substantial share of features in any untargeted metabolomics dataset stays unannotated. Whether that unannotated residue represents undiscovered biological chemistry or measurement artifact, particularly ions formed by fragmentation during electrospray ionization, has become an active point of disagreement among leading metabolomics groups, as laid out in a recent unannotated signals debate that examines how much of the so-called dark metabolome is chemistry versus instrumentation.
The practical implication for working scientists is caution rather than dismissal in either direction. Treating every unannotated feature as a hidden novel metabolite risks chasing artifacts, while dismissing the entire unannotated fraction as noise risks overlooking genuine biology, and current AI annotation tools cannot resolve that ambiguity on their own for any single feature.
Resolving individual unannotated features still depends on targeted follow-up work, including accurate mass measurement, isotope pattern analysis, and, where feasible, comparison against a synthesized or purchased reference standard. AI narrows the search space considerably and prioritizes which unannotated features are worth that follow-up effort, but it does not remove the need for the confirmatory chemistry itself.
AI Untargeted Metabolomics LC-MS Tools Still Need a Scientist in the Loop
AI has made untargeted metabolomics tractable at a scale that manual annotation could never reach, compressing feature detection, spectral matching, and statistical triage into workflows that a single analyst can now run against thousands of samples in a study. What it has not done is solve annotation itself, since most features detected in any given experiment still lack a confirmed chemical structure by the end of the pipeline.
For analytical scientists working with LC-MS metabolomics data, the realistic use of AI is as a triage and prioritization layer rather than a final answer. Confidence levels, orthogonal confirmation, and an honest accounting of the dark metabolome remain part of the job long after the algorithm has finished ranking its candidates, a discipline that carries over to the other emerging separation science applications in food safety, environmental, and clinical analytical chemistry that face the same annotation ceiling. Treated this way, AI earns its place in the untargeted metabolomics workflow without being asked to do more than it can currently deliver.
This article was produced under Separation Science's AI Editorial Guidelines.
