Running untargeted spatial metabolomics differs from a targeted assay in a way that matters more than acquisition settings suggest. In targeted work, you know what you are looking for, so a wrong answer often looks wrong. In untargeted work, every detected feature is a candidate finding, which means an artefact introduced before the sample reached the instrument is indistinguishable from biology — and the published evidence on that is more alarming than most workflow guidance admits.
Key Takeaways
|
Targeted or Untargeted? What Changes in Space
In solution-phase metabolomics, the distinction is familiar. In imaging it shifts, because as covered in Spatial Metabolomics and Lipidomics by Mass Spectrometry Imaging, imaging acquisition is inherently untargeted: the instrument records everything that ionises at each position, whether you intended it or not. What you choose is not whether to be untargeted at acquisition but whether to be targeted at interpretation.
| Targeted Interpretation | Untargeted Interpretation |
You know in advance | Which species you will report, and its identity | Neither |
Annotation burden | Low. Identity is established beforehand | The dominant difficulty of the experiment |
Error visibility | Reasonable. An implausible result for a known analyte looks implausible | Poor. Any feature is a potential finding, including artefacts |
Optimisation objective | Sensitivity for the target | Breadth of coverage across chemical classes |
Pixel size logic | As fine as the target signal supports | Coarser, since coverage falls as pixels shrink |
Validation route | Standards and internal standards for the target | Orthogonal confirmation of selected findings after the fact |
Table 1. Targeted against untargeted interpretation of imaging data. The error visibility row drives everything else in this article.
Working in analytical science?
Register for a FREE Separation Science account to subscribe to the Separation Science Newsletter.
That third row is the crux. A targeted assay has a built-in plausibility check, because you hold expectations for a known analyte. An untargeted experiment does not, so anything that systematically alters the metabolome before acquisition will be reported as a finding rather than caught as an error. Which is why this article treats sample handling before it treats acquisition.
Why Does Sample Handling Dominate Untargeted Results?
Because metabolites turn over on timescales comparable to how long it takes to get tissue from a subject into liquid nitrogen. Transcripts and proteins are comparatively stable across the minutes involved in dissection. Energy metabolites are not.
The Finding That Should Change Your Protocol A 2024 PLOS One study of cryopreservation timing in human liver tissue collected normal tissue at three points during liver resection and analysed it by NMR and LC-MS. Time at cryopreservation was the principal variable contributing to differences between specimen metabolomes, superseding even interindividual variability. Succinate, alanine, glutamine, arginine, leucine, glycerol-3-phosphate, lactate, AMP, glutathione, and NADP increased with delay, while aspartate, citrate and isocitrate, ADP, and ATP decreased. Read that in the context of an untargeted study design. If handling time contributes more variance than the difference between patients, then a comparison between two groups whose samples were collected under different logistics is measuring the logistics. And because the experiment is untargeted, nothing in the data will tell you that. |
The speed involved is documented precisely. A 2022 study in Molecular Metabolism on mouse tissue harvest-induced hypoxia dissected the left lateral lobe of liver from anaesthetised mice and froze it by liquid nitrogen-temperature freeze clamping immediately, at about one second, and after 30-second, 1-minute, 3-minute, and 10-minute delays. Compared with immediate freezing, 31 metabolites had significantly increased by 30 seconds and 128 by 10 minutes. Averaging across the broad metabolome, the authors calculated a time to one-half fold-change of approximately 3.58 plus or minus 0.32 minutes.
That last figure is the one worth carrying away: the broad metabolome reaches half of its total measured change in under four minutes. The same work found that delay induced both false-negative and false-positive between-genotype differences when comparing wildtype against liver-specific mitochondrial pyruvate carrier knockout mice, and that carbon-13 isotopologue abundances and enrichment percentages shifted as well, with the succinate m+3 isotopologue increasing many-fold in a way consistent with TCA cycle reversal. So delay does not merely scale metabolite levels — it restructures apparent flux.
Published timescales nonetheless vary enormously by tissue, which is the practically important part.
System | Delay Examined | Effect Reported | Practical Reading |
Mouse liver | Immediate (approx. 1 second) to 10 minutes | 31 metabolites significantly increased by 30 seconds and 128 by 10 minutes; half-maximal metabolome change in approx. 3.58 minutes; false-negative and false-positive between-genotype differences induced | Sub-minute matters. Purine nucleotide degradation products are a high dynamic range marker of delay |
Human liver | Across a surgical resection workflow | Cryopreservation timing was the principal source of variance between specimens, above interindividual variability | Clinical logistics must be standardised, not merely recorded |
Breast cancer xenograft | 0 to 120 minutes | No significant change in individual metabolites within 30 minutes; thereafter choline rose while ascorbate, creatine, and glutathione fell | Some tissues tolerate tens of minutes. Do not assume yours is one |
Rodent brain | Seconds | Even with immersion in liquid nitrogen within one second of cervical dislocation, ATP and phosphocreatine were reduced two- and three-fold relative to optimally frozen cortex | Brain energy metabolites may be unrecoverable by freezing alone |
Table 2. Published freezing-delay effects across four systems. Acceptable delay differs by more than two orders of magnitude between them, so it must be established for your tissue rather than adopted from a protocol.
The brain row deserves expanding, because it contains an uncomfortable point about standard practice. A 2020 analysis of labile brain metabolite measurement and harvest procedures argues that unequal, very low regional ATP levels and low ATP/ADP ratios reported in imaging MS studies are explained by rapid metabolism during postmortem ischaemia rather than by biology. It states that procedures which do not prevent postmortem autolysis, explicitly including decapitation, brain removal and dissection, and snap freezing, are commonly used, and that the requirement for inactivation of enzymes by freezing or heating is not widely recognised outside the neurochemistry discipline.
That is a direct challenge to imaging results many readers will have seen published. The practical response is to treat snap freezing as a starting point rather than a solution for labile species, and to consider heat stabilisation where the analytes warrant it. A comparison of heat stabilisation against conventional fresh freezing for MALDI imaging of brain tissue found that stabilisation left dopamine, norepinephrine, and serotonin levels unchanged while significantly lowering their downstream metabolites, indicating reduced postmortem turnover, and enabled detection of an increased number and higher levels of neuropeptides. Several low-abundance neuropeptides remained intact and were exclusively imaged in heat-stabilised brains, whereas degradation fragments appeared in the fresh frozen tissue. The wider question of standardising collection procedures for metabolomic work is treated in the brain collection standards literature.
Four handling practices follow, and none is difficult.
- Record time to freezing for every specimen, not as a nicety but as a covariate. If you cannot standardise it, you at least need to be able to test whether it explains your findings.
- Standardise the route, not just the endpoint. Where clinical collection is involved, the documented practical fix is placing collection equipment as close to the operating theatre as possible and defining the handling procedure for surgical staff.
- Establish your own tolerance. Run a short delay series on your tissue type, freezing replicate pieces at several intervals, and identify which features move. Purine nucleotide degradation products are a documented high dynamic range marker of delay.
- Balance handling across comparison groups. If delay cannot be eliminated, distribute it evenly across the groups you intend to compare, so it cannot masquerade as your result.
Acquisition for Discovery
Discovery acquisition optimises for breadth, and that inverts several instincts carried over from targeted work.
- Choose coarser pixels than feels ambitious. Coverage narrows as pixels shrink, because the species surviving smaller sampling volumes are the abundant and readily ionised ones. The arithmetic is in Spatial Resolution vs. Sensitivity in MS Imaging: The Fundamental Trade-off.
- Run both ion modes. Positive and negative mode return substantially different species lists, so a single-mode discovery experiment has surveyed part of the metabolome. Use adjacent sections rather than compromising conditions.
- Prioritise mass resolving power. With no separation before ionisation, resolving power distinguishes near-isobaric species, and in the low mass range it also helps separate analyte signal from matrix-derived ions, as discussed in MALDI Imaging Mass Spectrometry: How It Works.
- Expect to run more than one preparation. No single matrix or solvent system extracts the whole metabolome. Broad coverage usually means several conditions on adjacent sections rather than one optimised condition.
One acquisition decision specific to discovery is worth flagging: include a region you expect to be uninteresting. Homogeneous background tissue provides an internal reference for what a null distribution looks like in your own data, which is genuinely useful when every feature is a candidate. Without it, distinguishing a real localisation from a preparation artefact is harder than it needs to be.
Feature Detection and Alignment
Processing choices at this stage determine which features exist to be interpreted, and they are frequently made by software defaults rather than by decision. Four matter most.
Step | What the Choice Affects | Discovery-Mode Guidance |
Peak picking and binning | Which signals become features at all | Set tolerance from your measured mass accuracy, not a default. Too wide merges distinct species; too narrow splits one |
Mass alignment across pixels | Whether the same species is recognised as one feature throughout the image | Essential on FT instruments where pixel-to-pixel mass shifting can occur. Use lock mass during acquisition where available |
Normalisation | Apparent intensity relationships between regions | Try more than one approach and check whether conclusions survive. Normalisation can create structure as well as remove it |
Feature filtering | How many candidates reach interpretation | Filter on spatial coherence rather than intensity alone, since a real localisation is spatially structured and noise usually is not |
Table 3. Processing decisions that determine which features you interpret. Software defaults are a starting point, not a method.
On data volume, an untargeted imaging experiment produces a hyperspectral cube rather than a peak list, and at fine rasters that becomes large enough to constrain how you work rather than merely where you store it. Because that is a shared problem across this cluster rather than one specific to metabolomics, it is treated in Analyzing Mass Spectrometry Imaging Data: Processing, Statistics, and Multimodal Integration along with software options and computational approaches.
How Do You Interpret an Untargeted Map?
With more discipline than a targeted result requires, because the absence of a target list removes the constraint that ordinarily prevents over-interpretation. Four questions to ask of any candidate finding, in order.
- Is the localisation spatially coherent? A real distribution has structure corresponding to something. Speckle, edge-concentrated signal, or patterns aligned with the acquisition raster suggest artefact rather than biology.
- Does it correspond to histology? Co-registration with a stained section is the strongest available sanity check. A feature tracking a recognisable structure is more credible than one tracking nothing.
- What evidence level supports the annotation? Accurate mass gives candidate formulas, not molecules. The confidence ladder is set out in the section hub, and databases and reporting practice in Metabolite Annotation and Databases for MS Imaging.
- Could handling explain it? Given Table 2, this belongs on every checklist. Features associated with energy metabolism and purine degradation warrant particular suspicion if handling was not tightly controlled.
Then confirm rather than conclude. The productive pattern is discovery followed by orthogonal validation: use the image to identify which species and regions merit attention, then confirm the specific findings that matter by on-tissue fragmentation, ion mobility, or extraction of the region followed by a conventional separation-based workflow. Imaging is unusually good at telling you where to look and comparatively weak at telling you exactly what you are looking at, and a workflow respecting that division performs better than one asking imaging to do both.
What Does a Realistic Workflow Look Like?
In sequence, with the dependency order that matters. Optimising a later stage before an earlier one is wasted effort, because each stage constrains the next.
- Fix handling first. Establish and document time to freezing, standardise the collection route, and balance any residual variation across comparison groups. Nothing downstream repairs this.
- Settle sectioning and mounting. Thickness and mounting affect signal and delocalisation. Note that washing steps used in peptide and protein workflows will also remove water-soluble metabolites, so protocols cannot be borrowed across analyte classes without checking.
- Choose ion mode and preparation for coverage. Test candidate matrices or solvent systems on your tissue and count detected features in both polarities. This is where coverage is won or lost.
- Set pixel size from the question, then check the time budget. For discovery, favour coverage. Confirm the acquisition is feasible before committing a cohort to it.
- Establish processing parameters on a pilot section. Peak picking tolerance, alignment, and normalisation, decided deliberately and then held constant.
- Acquire the cohort with handling and processing held fixed. Changing parameters mid-study creates a batch effect no analysis will separate from biology.
- Annotate with evidence levels recorded, then validate selected findings orthogonally.
The recurring theme is that untargeted spatial metabolomics rewards decisions made in advance and punishes decisions discovered mid-study. That is true of most analytical work and more true here, because the absence of a target list means the data itself will not object. For where this sits alongside the other spatial modalities, see Spatial Analysis in Analytical Science: Mass Spectrometry Imaging and Spatial Omics.
This article was produced under Separation Science's AI Editorial Guidelines.
