Doing metabolite annotation in MS imaging well begins with accepting a constraint that is rarely stated directly: the community framework for identification confidence was built around chromatography, and imaging has none. That is not a gap awaiting better instruments. It is a structural feature of the technique, and understanding it changes what you should claim and how you should spend your effort.
Key Takeaways
|
Why Annotation Is Hard in MS Imaging
Three difficulties compound, and only the first is shared with conventional metabolomics.
- Mass does not determine identity. Isomers share a molecular formula exactly, so no improvement in mass accuracy separates them. This is true of all mass spectrometry.
- There is no separation before ionisation. Everything present at a pixel is ionised together, which both crowds the spectrum and removes the separation dimension that conventional workflows rely on, as set out in Mass Spectrometry Imaging: Principles, Techniques, and Applications.
- Fragmentation is expensive. On-tissue tandem MS is possible but costs acquisition time at every pixel where it is applied, so it is typically reserved for selected targets rather than applied across an image.
The consequence is that most imaging annotations rest on MS1 evidence alone, which is a weaker position than a routine LC-MS experiment occupies. The scale of the resulting gap is documented: as discussed in Spatial Metabolomics and Lipidomics by Mass Spectrometry Imaging, the developers of the field’s standard annotation engine describe the unassignable majority of imaging data as dark matter.
One terminological warning before going further. In this literature, the abbreviation MSI means both mass spectrometry imaging and the Metabolomics Standards Initiative, and both appear in discussions of annotation confidence. This article spells out the standards body in full throughout to avoid the collision.
Working in analytical science?
Register for a FREE Separation Science account to subscribe to the Separation Science Newsletter.
Why Can’t MS Imaging Reach Level 1?
Because level 1 was defined to require something imaging does not produce. This is worth working through carefully, since it determines what you can honestly claim.
The Metabolomics Standards Initiative originally proposed a four-tier system for communicating identification confidence, which was subsequently refined by Schymanski and colleagues into a five-level framework tailored to the capabilities of high-resolution mass spectrometry. The levels run from level 5, exact mass alone, to level 1, a confirmed structure supported by a reference standard, incorporating molecular formula, spectral library matches, and structural inference along the way.
The top of that scale has a specific requirement. Level 1 is reserved for compounds confirmed by direct comparison with an authentic standard measured under identical analytical conditions, and the standards initiative definition requires a minimum of two independent and orthogonal pieces of data from that standard. In practice, for LC-MS work, those axes are accurate mass, fragmentation, and chromatographic retention time.
Level | What It Represents | Evidence Required | Available in Imaging? |
5 | Exact mass only | Accurate mass measurement | Yes. This is the baseline for every imaging feature |
4 | Unequivocal molecular formula | Accurate mass plus isotope pattern | Yes, with high mass resolving power |
3 | Tentative candidate or compound class | Formula plus partial structural evidence | Yes, with on-tissue fragmentation or class-specific ions |
2 | Probable structure by library or diagnostic evidence | Spectral library match or strong structural inference | Partly. Library matching is harder without a separation dimension |
1 | Confirmed structure | Authentic standard under identical conditions, conventionally including retention time | No. There is no chromatography, so no retention time to match |
Table 1. The five-level confidence framework mapped against what mass spectrometry imaging can supply. The framework is as published; the availability column is our assessment of what imaging can and cannot provide.
This Is a Structural Limit, Not a Maturity Problem Coverage of imaging annotation often implies the field will catch up with LC-MS metabolomics as instruments improve. On this particular axis it will not, because the limitation is not sensitivity or resolving power but the absence of a separation stage. A technique with no chromatography cannot produce a retention time to match against a standard, however good its mass analyser. Two useful conclusions follow. First, report honestly: an imaging annotation is normally a level 4 or level 3 claim, and describing it as an identification overstates it. Second, spend effort where it can actually move you — which means the axes imaging can still add, rather than the one it cannot. |
Databases and Reference Resources
Database choice is the most consequential decision in an annotation workflow and the least deliberated. It sets both what can be found and how many candidates each finding carries.
Resource | Scope | How It Behaves in Imaging Annotation |
HMDB, Human Metabolome Database | Broad coverage of human metabolites, including many not expected in a given tissue | Comprehensive, and therefore returns large candidate lists per annotated ion |
LIPID MAPS | Lipid classification, nomenclature, and structures | Essential for lipid work, and supplies the notation that encodes structural evidence level |
Expert-curated core databases | Deliberately restricted to metabolites plausibly present and detectable | Fewer candidates per ion and better precision; the METASPACE developers report improved results over general databases |
Pathway and compound registries | Biochemical context and cross-references | Useful for interpretation rather than assignment |
Specialised community databases | Domain-specific compounds, for example microbial specialised metabolites | Substantially improves annotation in contexts general databases cover poorly |
Table 2. Reference resources for imaging annotation. The trade-off in the third column is the practical point: comprehensiveness and precision pull against each other.
The counterintuitive guidance is to prefer the smaller database. Because imaging annotation works from MS1 evidence, every additional structure in the database that shares a molecular formula with a real feature becomes another candidate you cannot exclude. A comprehensive database therefore inflates ambiguity without adding information. The METASPACE machine learning work reports exactly this: introducing an expert-curated core database improved results relative to general databases, and the model itself was trained and evaluated across 1,710 datasets from 159 researchers at 47 laboratories, spanning animal and plant contexts.
For lipids specifically, the LIPID MAPS classification and shorthand notation standard does double duty: it is a structural reference, and it provides the notation that declares how much structure you have actually established, which is developed in Spatial Lipidomics in Tissue.
How Does FDR-Controlled Annotation Work?
By asking how often the method would produce an annotation that cannot be real, and using that rate to calibrate confidence in the ones that might be.
The approach implemented for imaging generates implausible decoy ions, scores real candidate annotations against them, and reports annotations at a stated false discovery rate by that ranking. Scoring combines several measures, including agreement between measured and theoretical isotope patterns and the spatial coherence of the ion image, on the reasoning that a real molecular distribution is spatially structured while noise generally is not.
Three things follow that are worth understanding before quoting an FDR figure.
- The FDR is conditional on the database. It expresses the rate of false formula assignments given the search space you chose. Changing the database changes the figure, so an FDR without a stated database is uninterpretable.
- It controls formula assignment, not structure. An annotation passing a 10 percent FDR threshold is a formula reported at that confidence, still potentially corresponding to many isomeric structures.
- Spatial coherence is doing real work. This is a genuine advantage imaging has over infusion experiments: the image itself is evidence, because a spatially random distribution is unlikely to be a real metabolite localisation.
That third point deserves emphasis because it partly offsets the missing retention time. Imaging gains an evidence type that conventional metabolomics does not have, namely the spatial structure of the signal. It is not equivalent to a separation dimension, since it speaks to whether a signal is real rather than to what it is, but it is not nothing.
Isomers and In-Source Fragments
Two categories of ambiguity behave differently and require different responses. Isomers are a structural problem; in-source fragments are an artefact problem, and the second is more often overlooked.
In-source fragmentation occurs when a molecule breaks apart during desorption and ionisation rather than in a collision cell. The resulting fragment appears in the MS1 spectrum as though it were an independent species, and it will be annotated as one if its mass matches a database entry. In imaging, this is particularly troublesome because there is no chromatographic separation to reveal that the fragment and its parent share an origin.
Three practical defences, in order of usefulness.
- Check spatial co-localisation. A fragment will map exactly onto its parent, because it is generated from it at the point of desorption. Perfect co-localisation between a smaller and a larger species related by a plausible neutral loss is a strong indication of in-source fragmentation rather than two co-regulated metabolites.
- Look for characteristic neutral losses. Losses of water, phosphate, or head groups relate fragments to parents predictably within a class, so they can be anticipated.
- Moderate the ionisation energy. Reducing laser fluence lowers in-source fragmentation, at some cost in signal, which is a trade worth testing during method development.
The first of those is the one to internalise, because it inverts an instinct. Two ion images that overlay perfectly look like a strong biological result and are frequently an artefact of one molecule being counted twice. Suspiciously exact co-localisation warrants checking rather than celebration.
How Do You Improve Confidence Without Retention Time?
By adding the axes that remain available. Given the structural argument above, this is where method development effort actually pays, and one option matters more than the others.
Approach | What It Adds | Practical Cost |
High mass resolving power | Confident molecular formula assignment, reaching level 4 | Acquisition speed, which matters given pixel counts |
Ion mobility and collision cross-section | A genuinely orthogonal structural descriptor, and the closest available substitute for retention time | Modest. Separation is achieved without appreciably slowing the raster |
On-tissue tandem MS | Structural evidence toward level 3 or 2 | Acquisition time, so usually applied to selected targets or regions |
Curated database restriction | Fewer candidates per annotation, improving precision | Risk of excluding a genuinely present but unlisted compound |
Spatial coherence filtering | Confidence that a feature is real rather than noise | Little, and it is imaging-specific evidence worth exploiting |
Orthogonal validation by extraction | Can reach level 1 for selected compounds, since the extract can be chromatographed | Destroys spatial information for that sample, so it is a separate experiment |
Table 3. Routes to higher annotation confidence in imaging. The second and last rows are the two most strategically important.
Ion mobility deserves the emphasis for a specific reason. Conventional metabolomics has accurate mass, fragmentation, and retention time as orthogonal axes, with collision cross-section increasingly added as a further one. Imaging permanently lacks retention time, so cross-section is not an incremental refinement here but the replacement for a missing dimension. That is why it carries more weight in imaging than in LC-MS work, where retention time already does that job. A spatial metabolomics workflow using cyclic ion mobility with predicted collision cross-sections reported accuracy better than 0.4 percent relative to database values in multipass experiments, which was sufficient to improve filtering thresholds and resolve assignments that mass alone left ambiguous.
The final row of Table 3 is the honest route to a definitive identification, and it is worth stating explicitly: extract the region of interest and run a conventional separation-based experiment on it. That recovers retention time and can reach level 1, at the cost of the spatial information for that sample. Discovery by imaging followed by confirmation by extraction is not a workaround, but the appropriate division of labour, and it is the workflow pattern recommended in Untargeted Spatial Metabolomics Workflows.
Software, processing choices, and computational approaches to all of this are covered in Analyzing Mass Spectrometry Imaging Data: Processing, Statistics, and Multimodal Integration. For where annotation sits in the wider spatial landscape, see Spatial Analysis in Analytical Science: Mass Spectrometry Imaging and Spatial Omics.
This article was produced under Separation Science's AI Editorial Guidelines.
