AI tools for unknown compound identification in LC-MS have become essential to non-targeted analysis, the branch of mass spectrometry where a sample is screened broadly rather than tested against a predefined list of known compounds. Environmental contaminant discovery, food fraud detection, metabolomics, and forensic toxicology all rely on identifying compounds that were never specifically targeted, which means a substantial fraction of the features detected in any non-targeted run have no obvious answer waiting in a reference library. AI tools are meaningfully improving how many of those unknowns get annotated, but the core analytical challenge, distinguishing a confident identification from a plausible guess, remains squarely the scientist's to manage.
This article is part of Separation Science's coverage of machine learning in mass spectrometry; the guide to machine learning for MS and spectral data analysis covers the wider landscape, and the broader AI in analytical science overview sets the full context. For a researcher's perspective on the proteomics side of this challenge, the Q&A with Talus Bio's Will Fondrie is a useful companion read.
Key Takeaways
|
Working in analytical science?
Register for a FREE Separation Science account to subscribe to the Separation Science Newsletter.
Why Unknown Compound Identification Is Hard
Targeted analysis asks a manageable question: is this specific compound present, and at what concentration? Non-targeted analysis asks a much harder one: what is present in this sample at all? A single LC-MS run in non-targeted mode routinely detects thousands of features, mass-retention time pairs that represent real chemical signals, and for the large majority of them there is no predefined expectation of what they should be.
The scale of the problem comes from several compounding factors:
- Chemical space vastly exceeds library coverage. Reference spectral libraries, however extensive, cover a small fraction of chemically plausible structures. PubChem alone lists well over 100 million compounds; even the largest curated MS/MS spectral libraries cover a tiny fraction of that space.
- Isomers and structurally similar compounds confound mass-based matching. Many distinct compounds share the same molecular formula or extremely similar fragmentation behaviour, so accurate mass and even fragmentation pattern alone often cannot uniquely distinguish them.
- Transformation products and unknowns have no entry anywhere. Metabolites, environmental degradation products, and genuinely novel compounds frequently have no measured reference spectrum in any database, public or commercial.
- Manual interpretation does not scale. Even an experienced analyst evaluating fragmentation patterns and chemical plausibility by hand can realistically work through a small number of priority features, not the thousands a single non-targeted run produces.
Non-targeted analysis routinely detects thousands of features per sample. The honest starting point is that most of them will not be identified with certainty, and AI changes how many, not whether that gap exists.
How Spectral Library Matching Works (and Its Limits)
Spectral library matching is the foundational, most reliable approach to compound identification, and it remains the first step in almost every workflow regardless of how much AI assistance follows. The principle is straightforward: compare the experimental spectrum, typically the fragmentation pattern from MS/MS, against a database of reference spectra measured from known compounds, and rank candidates by similarity.
What library matching does well, and where it stops:
- High confidence when a true match exists. If the compound and a closely matching reference spectrum, ideally measured on similar instrumentation, both exist in the library, matching is fast and reliable, often reaching the highest confidence tiers without further work.
- Sensitivity to instrument and acquisition differences. Fragmentation patterns vary meaningfully with collision energy, instrument type, and ionisation conditions, so a poor match score does not always mean the compound is absent from the library, only that the conditions differ.
- No answer for anything not in the library. This is the fundamental limit: library matching can only confirm what someone has already measured and catalogued. For a genuinely novel compound, a transformation product, or simply an under-represented chemical class, there is no entry to match against, however good the matching algorithm is.
This limit is exactly where AI-enhanced approaches earn their place. Library matching answers the question of whether a compound has been seen before; AI-assisted in silico methods attempt to answer the much harder question of what a never-before-seen compound might be.
AI-Enhanced In Silico Fragmentation Prediction
In silico fragmentation prediction sidesteps the library-matching limit entirely: instead of comparing an experimental spectrum to a measured reference, it predicts what fragmentation pattern a candidate structure should produce and compares that prediction to the experimental data. This means candidates can be scored even when no one has ever measured a reference spectrum for them, which is the single most important capability for genuinely novel compound annotation.
The clearest demonstration of this approach is CSI:FingerID, developed by the Böcker group, which combines fragmentation tree computation with machine learning to predict a molecular fingerprint from a tandem mass spectrum, then searches structure databases for the best match to that predicted fingerprint. In large-scale evaluation, the method found 150 percent more correct identifications than the next-best search approach available at the time, and 5.4-fold more unique identifications than competing methods, a substantial and well-documented improvement.
How this class of tool changes the identification workflow:
- Candidates are scored on predicted chemistry, not just observed similarity. Because the method reasons from a fragmentation tree and predicted molecular fingerprint rather than spectrum-to-spectrum comparison, it can meaningfully rank candidates that have never been measured.
- Performance improves with more training data. As is typical of machine learning approaches, evaluation work indicates further performance gains as more reference training data becomes available, which has continued as spectral databases have grown.
- Database search still matters. These methods score candidates drawn from large structure databases such as PubChem; the prediction step narrows an enormous candidate pool to a ranked shortlist, it does not eliminate the database search step entirely.
The practical implication is that in silico fragmentation prediction has moved from a research curiosity to a standard part of serious non-targeted workflows, particularly in environmental and metabolomics contexts where novel compound annotation is the norm rather than the exception.
Machine Learning for Novel Compound Annotation
Beyond fragmentation prediction, several complementary machine learning approaches contribute to annotating compounds with no library entry, each addressing a different piece of the puzzle. Retention time prediction, of the kind demonstrated at scale by the METLIN SMRT dataset, helps rule out candidates that are structurally plausible from mass and fragmentation alone but chromatographically inconsistent with the observed elution time, narrowing the field further.
Other approaches contributing to the novel-annotation problem:
- Compound class prediction. Tools that predict the broad chemical class of a feature, without requiring a full structural match, provide useful triage information even when a specific structure cannot be confidently assigned, helping prioritise which unknowns deserve further investigation.
- Molecular network and similarity approaches. Methods that cluster spectra by similarity can propagate identifications from a known compound to structurally related unknowns in the same cluster, a particularly useful approach in natural products and metabolomics research.
- Biotransformation and metabolite prediction. For suspected metabolites or transformation products, tools that predict likely biotransformation products from a parent compound can generate candidate structures to test against the observed spectrum, addressing cases where the parent is known but the product is not.
None of these approaches, alone or combined, eliminates the fundamental gap between detection and confirmed identification. They shift the odds meaningfully in the analyst's favour, converting an otherwise unannotated feature into a ranked, testable hypothesis rather than a dead end.
Confidence Level Frameworks for Unknown Identification
Because AI-assisted identification produces a ranked candidate with a confidence score, not a certainty, the field has developed structured frameworks for communicating exactly how confident a given identification actually is. The most widely adopted is the five-level scheme proposed by Schymanski and colleagues in 2014, which has been cited in thousands of subsequent studies and has become close to a field standard, particularly in environmental and non-targeted screening.
Level | Designation | What Is Required |
1 | Confirmed structure | Reference standard measured under the same conditions, matching retention time and spectrum |
2 | Probable structure | Library spectrum match or diagnostic evidence strong enough to assign a structure with reasonable confidence, but no reference standard run |
3 | Tentative candidate | Structural information available but insufficient to distinguish among multiple isomers or close candidates |
4 | Unequivocal molecular formula | Formula established with confidence from accurate mass and isotope pattern; no structural assignment possible |
5 | Exact mass of interest | Only the mass is known; no structural or formula-level information beyond that |
AI-assisted tools typically help move a feature from level 5 toward level 2 or 3: they rarely substitute for the reference-standard confirmation that level 1 requires. Reporting the appropriate confidence level alongside any AI-assisted identification, rather than presenting a ranked candidate as a confirmed result, is the single most important discipline in using these tools responsibly.
A model's confidence score and the actual correctness of an identification are not the same thing. Confidence-level frameworks exist because that distinction has real consequences for how a result can be used.
Practical Tools: SIRIUS, CSI:FingerID, and MetFrag
Several open-source and academic tools have become standard reference points for AI-assisted unknown compound identification, and understanding what each one specifically does is more useful than treating them as interchangeable. SIRIUS, developed by the Böcker group, is the umbrella framework that combines isotope pattern analysis, fragmentation tree computation, and several integrated modules, including CSI:FingerID for structure database searching and CANOPUS for compound class prediction when no structural database match exists at all.
How the main tools in this space relate to each other:
- SIRIUS. The overall framework: takes MS1 isotope data and MS2 fragmentation data, computes the most probable molecular formula, and constructs a fragmentation tree that becomes the input for downstream structure identification.
- CSI:FingerID. Operates within SIRIUS to predict a molecular fingerprint from the fragmentation tree and search structure databases for the best match, the step responsible for the substantial identification gains described above.
- CANOPUS. Predicts the broad compound class directly from the predicted molecular fingerprint, without requiring any database search, providing useful triage information even for compounds with no structural database match at all.
- MetFrag. A widely used, independently developed in silico fragmentation tool that scores candidate structures from chemical databases against observed fragmentation patterns, frequently used alongside or as an alternative to the SIRIUS suite.
These tools are free for academic use, well documented, and have a substantial track record in independent benchmarking exercises (CASMI, the Critical Assessment of Small Molecule Identification), which is a meaningful signal of reliability in a field where vendor claims can be difficult to verify independently. Most non-targeted workflows now use several of these tools in combination, rather than relying on any single method, precisely because each contributes a different kind of evidence to the identification.
What This Means for Your LabTreat AI-assisted unknown compound identification as a powerful triage and hypothesis-generation tool, not a substitute for confirmation. Start with spectral library matching, since it remains the most reliable route when a true match exists, then layer in silico fragmentation prediction and complementary tools for everything the library cannot answer. Report the appropriate confidence level for every identification rather than presenting a ranked candidate as settled fact, and treat reference standard confirmation as the bar for any result that will inform a regulatory, clinical, or otherwise consequential decision. Done this way, AI substantially raises the proportion of features that get a useful answer, without pretending the gap between detection and certainty has closed. For the proteomics side of this same challenge, the Q&A with Talus Bio's Will Fondrie offers a complementary perspective, and the guide to machine learning for MS and spectral data analysis sets the wider context. |
This article was produced under Separation Science’s AI Editorial Guidelines
