Mass Spectrometry Imaging: MSI Workflow

The instrument is rarely what decides whether a mass spectrometry imaging (MSI) experiment succeeds. Sectioning, mounting and matrix application decide it. By the time the laser or spray fires, the spatial information has either survived sample preparation or it has not, and no amount of downstream processing recovers a molecule that migrated 200 µm during a wet matrix coat. The second thing that sinks MSI projects happens at the other end: reading a formula-level annotation as if it were a confirmed structure, or reporting a pattern that normalization created rather than revealed.

This guide walks the workflow in the order the decisions bind, and flags the point in each step where the outcome is actually determined.

What MSI measures, and what it does not

MSI acquires a complete mass spectrum at every position in a raster across a tissue section. The result is a data cube: two spatial axes and one m/z axis. Selecting any m/z value renders an ion image showing where that species was detected and at what relative intensity. Because the readout is the mass spectrum itself, MSI is label-free and untargeted — you do not need to know what you are looking for in advance, and you do not need an antibody or a probe. That is the fundamental contrast with immunohistochemistry, where you only see what you chose to stain for.

Three limits follow directly from how the measurement works, and all three are routinely overstated in published figures:

Intensity is not concentration. Ionization efficiency varies with local chemistry, so the same molecule at the same concentration produces different signal in gray matter than in white matter.
Annotation is usually at the level of the molecular formula, not the structure. Isomers with identical elemental composition are one peak.
Every pixel is a destructive, one-shot measurement. You get one acquisition per section per polarity unless the method explicitly supports reapplication or a second modality.

Choose the ionization source from the resolution you need and the analytes you must see

These two requirements trade against each other, and the trade is the first real decision in the project.

MALDI

Matrix-assisted laser desorption/ionization is the workhorse. It offers the most favourable balance of preparation effort, chemical sensitivity and spatial resolution, and it covers lipids, metabolites, peptides, glycans and drugs. Reported resolution reaches roughly 1.4 µm with atmospheric-pressure sources, but most published cancer metabolomics work runs at around 10 µm, and typical untargeted small-molecule acquisition covers about m/z 80–1000. The cost is that MALDI needs a matrix, and matrix cluster ions crowd the low-mass region, which makes species below roughly 600 Da harder to detect cleanly.

DESI and other ambient sources

Desorption electrospray ionization sprays charged solvent at the tissue surface at ambient pressure, so it needs no matrix and does relatively little damage to the section. The theoretical resolution floor is about 10–20 µm, but most published studies run at 50–200 µm. Nano-DESI reaches around 10 µm; AFADESI runs around 100 µm with high untargeted coverage but is sensitive to parameter drift. Choose DESI when matrix-free preparation matters more than resolution, or when you want the section back afterwards.

SIMS and NanoSIMS

Secondary ion mass spectrometry pushes to the subcellular scale. NanoSIMS reaches around 50 nm lateral resolution and is the practical route to imaging stable-isotope-labelled lipids inside organelles — measuring ¹³C/¹²C ratio images rather than intact molecular species. Time-of-flight SIMS produces heavily fragmented secondary ions, which limits its usefulness for identifying intact lipids at subcellular scale. SIMS answers “where did this labelled atom go?” far better than “which lipid species is this?”

Fresh-frozen or FFPE: decide before you cut

Both MALDI and DESI accept fresh-frozen and formalin-fixed paraffin-embedded tissue, but they answer different questions. Fresh-frozen preserves small molecules, metabolites and intact lipids. FFPE cross-links proteins and washes out most small molecules, so FFPE MSI is effectively a spatial peptide experiment: you dewax, retrieve, digest on-tissue with trypsin, and image the resulting peptides.

A validated FFPE workflow gives a sense of how tightly the parameters need to be pinned down. One published optimisation used 5 µm sections on polylysine-precoated indium tin oxide (ITO) slides; dewaxing in two 2-minute xylene washes followed by a descending isopropanol series at 3 minutes each; antigen retrieval in 10 mM citric acid at 85 °C for 60 minutes, then 20 minutes at room temperature; trypsin at 50 nM in 20 mM ammonium bicarbonate applied as 15 sprayed layers at 5 µL/min; and 2,5-dihydroxybenzoic acid matrix at 0.1 mM in 0.1&percnt; TFA with 10&percnt; acetonitrile as seven layers at 7 µL/min. Acquisition ran at a 20 µm grid with a 10 µm laser spot over m/z 600–1500, with interday variability of 15.16&percnt;. Treat those figures as one worked example rather than a universal recipe — retrieval and digestion conditions are tissue- and instrument-dependent — but note the level of specification that reproducibility required.

Section thickness and mounting

The useful heuristic is that section thickness should approximate the width of one cell, commonly 8–20 µm, with 5–20 µm standard for small molecules in mammalian tissue. Plant tissue is different: around 50 µm is reported as a workable compromise between MSI performance and the practicality of cutting cell-wall-supported material.

Thicker is not safer. A section much thicker than a cell layer stacks distinct cell populations into one pixel and destroys the spatial contrast you ran the experiment to see. Thinner sections give cleaner spatial assignment but less material per pixel, which is exactly the sensitivity you are already short of at small pixel sizes.

Mount on conductive ITO-coated slides for vacuum MALDI. If you need a specific cell population rather than a full-section map, laser capture microdissection followed by conventional LC-MS is often the better instrument choice — deeper coverage, real quantification, at the price of losing the continuous spatial map.

Matrix selection is analyte- and polarity-specific

There is no default matrix. Published pairings that recur:

Lipids: DHB in positive mode; DAN (1,5-diaminonaphthalene) in negative mode for glycerolipids. In a six-matrix comparison on an AP-MALDI Orbitrap, THAP and CHCA detected the most lipid species in positive mode (155 and 136 respectively), while a DAN preparation in 70&percnt; acetonitrile led in negative mode with 136 lipids and performed best overall across both polarities.
Peptides below roughly 3000 Da: CHCA or DHB, positive mode.
Proteins above roughly 3000 Da: sinapinic acid, positive mode — with the caveat that proteins larger than about 25 kDa are not routinely detectable by MALDI-MSI.
Glucosinolates and similar acidic small molecules: 9-aminoacridine, negative mode.
Carbohydrates: DHB or CHCA; colloidal graphite substantially reduces matrix interference in the low-mass region.

If your project needs both polarities, plan for two sections or a method validated for dual-polarity acquisition. Running the wrong polarity for your analyte class is a common and completely avoidable loss of an entire experiment.

Laser post-ionization changes the sensitivity calculation

MALDI-2 fires a second laser roughly 400 µm above the ablation surface, intersecting the desorption plume after a set delay and ionizing neutrals that the primary event released but never charged. Reported gains exceed two orders of magnitude for some lipid classes — hexosylceramides, phosphatidylethanolamines, phosphatidylinositols, phosphatidylglycerols, phosphatidylserines, triacylglycerols and cholesterol esters — while phosphatidylcholines, already efficiently ionized, do not increase significantly. Matrix choice still matters under MALDI-2 and the ranking changes: norharmane gave the broadest neutral-lipid coverage in one four-matrix comparison, DHB the largest increase in total lipids detected. Do not assume a matrix optimised for conventional MALDI is optimal with post-ionization.

Matrix application: where spatial resolution is won or lost

This is the step most likely to quietly ruin the experiment, because a delocalized image still looks like an image.

Sublimation heats matrix under reduced pressure onto a cooled sample. It is solvent-free, so diffusion of most analytes during coating is almost eliminated, and it produces fine, uniform crystals at low cost. The trade-off is extraction efficiency: with no solvent, less analyte is drawn into the matrix layer.
Automated spraying gives better extraction and good reproducibility when the sprayer controls flow, temperature and pass geometry. Reported crystal sizes are small enough not to be the limiting factor — one DAN preparation sprayed with a robotic system produced roughly 1 µm crystals, supporting a confirmed 10 µm spatial resolution.
Manual airbrushing is where delocalization comes from. Inhomogeneous coating and lateral analyte migration occur easily and are operator-skill-dependent.
Sieving through a 20–50 µm mesh is fast and avoids delocalization entirely, but extraction efficiency drops.

The practical rule: solvent volume buys sensitivity and costs spatial fidelity. Every additional wet pass moves analytes. If you are targeting sub-20 µm pixels, sublimation or a well-characterised automated sprayer is not optional.

Set pixel size against the laser spot, not against the number you want to report

Pixel size and laser spot size are independent parameters, and conflating them produces two distinct errors. Setting a raster step smaller than the ablation spot without oversampling means adjacent pixels sample overlapping material, so the image is smoother than the data justifies. Setting the step much larger than the spot leaves unsampled tissue between pixels, which is legitimate but must be reported.

Sensitivity scales with the material in each pixel, so halving pixel size cuts the analyte available per spectrum roughly fourfold. Small-pixel MSI is therefore usually a sensitivity problem before it is an optics problem — which is precisely why post-ionization and high-field analysers matter more at 5 µm than at 50 µm. Choose the coarsest pixel that resolves the structure your biological question needs, and spend the recovered signal on mass resolving power instead.

Annotation: formula-level, FDR-controlled, isomer-blind

High-mass-resolution MSI supports database annotation at the level of the molecular sum formula. METASPACE, the community platform for this, computes a target–decoy false discovery rate by sampling decoy ions from implausible adduct combinations — 20 decoys per target formula in the machine-learning implementation — and reports annotations at fixed FDR thresholds of 5&percnt;, 10&percnt;, 20&percnt; and 50&percnt;. The reference collection behind the ML model spans 1,710 datasets from 159 researchers across 47 labs, and 84&percnt; of those datasets came from Orbitrap or FT-ICR instruments, which is a fair indication of the mass accuracy the approach assumes.

Two consequences for how you write up results:

An annotation is a formula, not a compound. The method cannot distinguish isomers; resolving them is explicitly named as future work requiring ion mobility. Writing “we detected [named lipid]” from a formula-level annotation alone overstates the evidence.
The FDR threshold is a reporting choice you must state. A 50&percnt; FDR list is a hypothesis-generating list, not a finding.

On-tissue tandem MS is the way to raise confidence. One AP-MALDI Orbitrap study identified 76 peaks by on-tissue MS/MS, demonstrating that lipid signal from tissue is sufficient for fragmentation-based confirmation. If a specific species carries the conclusion, fragment it. The same annotation-confidence logic applies in bulk workflows — see lipidomics by mass spectrometry for how extraction and acquisition choices constrain what can be identified.

Normalization: the step most likely to create a finding that is not there

Total ion current (TIC) normalization is the most widely used approach in MSI, and it is the default in most vendor software. Its failure mode follows from the arithmetic: dividing every pixel spectrum by its own total ion current means that a region genuinely richer in ionizable material has all of its individual features scaled down. The anatomical structure driving the TIC is thereby imprinted, inverted, onto every low-abundance ion image in the dataset. You then observe a “pattern” in a minor metabolite that is an artefact of the major lipids around it.

Alternatives in routine use include median fold change normalization, which removes biologically unrelated pixel-to-pixel variation using the median peak intensity per spectrum; intensity profile normalization, which applies a nonlinear, mass-dependent transformation per pixel and outperformed both TIC and MFC in a tumour-classification benchmark; sequential paired covariance; and normalization to an internal standard, which is standard practice in quantitative MSI. Related processing choices — peak picking by local maxima or Gaussian fitting, and mass alignment by correlation-optimised warping or known calibrant ions — also shape what the final images show.

The defensible practice is to inspect the TIC image itself before normalizing. If the TIC image already reproduces the anatomy, TIC normalization will contaminate every feature, and you should report results under more than one normalization to show the finding survives.

Quantification: MSI is semi-quantitative until you build calibration into the section

Ion intensity for the same concentration of a standard varies measurably across anatomical regions because local chemical composition changes ionization efficiency — a published study documented exactly this across gray and white matter in brain. Absolute quantification therefore requires putting known amounts of analyte through the same tissue-specific suppression the endogenous molecule experiences.

Sprayed standard addition is the current approach: apply stable-isotope-labelled analogues across the section by robotic sprayer rather than spotting them, because manual spots of around 1 mm diameter sit inhomogeneously across heterogeneous tissue. One protocol validated sprayer deposition gravimetrically at 0.092 ± 0.006 mg/cm² against a theoretical 0.095 mg/cm², achieved calibration linearity of R² = 0.995 and 0.999 for two neurotransmitters, and found agreement with HPLC-ECD reference measurements. Even so, the authors note that extraction efficiency still differs between endogenous analyte and externally applied standard. Report MSI values as relative unless you built and validated a calibration; a figure legend that says “concentration” without one is not supportable.

Co-registration with histology

MSI results become interpretable when overlaid on histology, and the alignment is a real analysis step, not a cosmetic one. Where the workflow permits, staining the same section after removing MALDI matrix (typically graded ethanol washes) gives the tightest correspondence because no physical deformation intervenes. Where a serial section is used instead, the two modalities differ in channel count (thousands of mass features versus up to four H&E channels), spatial resolution, aspect ratio and intensity distribution, and the sections carry elastic deformations relative to each other from handling.

Open tooling exists for this: MSIreg is an R package implementing a landmark-free co-registration workflow with quantitative accuracy metrics — Dice coefficient and normalized cross-correlation for global accuracy, and the Jacobian determinant of the displacement field to detect local expansion, contraction or folding. Reported case-study improvement ran from 82.88&percnt; to 95.20&percnt; Dice. Report a co-registration accuracy metric rather than asserting that the overlay looks right.

Data format, deposition and reporting

imzML is the common exchange format, developed on the HUPO-PSI mzML foundation with the controlled vocabulary extended for imaging-specific parameters such as x/y position and spatial resolution. It is a two-file format: an XML file (.imzML) holding metadata, and a binary file (.ibd) holding the spectra, linked by a shared UUID recorded in the imzML fileContent element. Two storage modes exist — continuous, where every spectrum shares one m/z array stored once, and processed, where each spectrum carries its own. The specification has been at version 1.1.1 since 2018. Converting to imzML at the end of acquisition is what frees the dataset from vendor software.

Reporting quality in the field is measurably poor. The MIAMSIE reporting template (Minimum Information About a Mass Spectrometry Imaging Experiment) defines 130 fields across five categories — general, sample preparation, MSI data, processing and other — reducing to about 81 for a typical experiment once conditional questions are applied, with the abbreviated MSIcheck covering 32 critical fields. When its authors applied MSIcheck to a sample of 19 MSI papers, none provided direct access to raw data, and calibration details, software versions, laser diameter and scan resolution were consistently under-reported. Those are precisely the fields a reader needs to reproduce or even interpret an ion image.

A minimum viable run plan

Fix the biological question, then the spatial scale it requires. That sets pixel size, and pixel size sets everything downstream.
Choose fresh-frozen or FFPE from the analyte class, not from what the biobank happens to hold.
Section at roughly one cell width; mount on ITO for vacuum MALDI; cut serial sections for histology and for the second polarity in the same sitting.
Select matrix and polarity for the analyte class; run a small test region before committing the section.
Apply matrix by sublimation or a controlled automated sprayer; record every application parameter.
Acquire with mass resolving power sufficient for formula-level annotation, and include on-tissue MS/MS for any species that carries a conclusion.
Inspect the TIC image before choosing a normalization; report the finding under at least two normalizations.
Annotate at a stated FDR threshold; describe results as formulas unless fragmentation confirmed the structure.
Co-register to histology and report an accuracy metric.
Export to imzML, deposit, and complete MSIcheck fields alongside the manuscript.

Frequently asked questions

What spatial resolution can mass spectrometry imaging actually achieve?

It depends entirely on the source. MALDI has been reported down to roughly 1.4 µm with atmospheric-pressure sources, though around 10 µm is typical in published work. DESI runs at 50–200 µm in most studies with a theoretical floor near 10–20 µm, and nano-DESI reaches about 10 µm. NanoSIMS reaches roughly 50 nm but images labelled atoms rather than intact molecules. Achievable resolution on your instrument is also constrained by matrix crystal size, section thickness and available signal per pixel, so treat vendor specifications as ceilings rather than expectations.

Can MSI identify specific molecules, or only masses?

Untargeted MSI annotates at the level of the molecular sum formula against a database, with a target–decoy false discovery rate. That cannot separate isomers, which share an elemental composition. To claim a specific structure you need on-tissue tandem MS, ion mobility separation, or orthogonal confirmation from an extract analysed by LC-MS.

Does MSI work on FFPE tissue?

Yes, but as a peptide experiment. Formalin fixation and paraffin processing cross-link proteins and remove most small molecules, so FFPE MSI workflows dewax, perform antigen retrieval, digest on-tissue with trypsin, and image the resulting peptides. Fresh-frozen tissue is required if you want intact lipids or metabolites.

Why do my ion images all look like the anatomy?

Most often, TIC normalization. If one region contains more total ionizable material, dividing each spectrum by its own TIC imprints that region’s outline into every feature, including ions with no biological relationship to it. Look at the unnormalized TIC image; if it already shows the structure you are seeing in your ion images, try median fold change, intensity profile normalization or an internal standard before believing the result.

Is mass spectrometry imaging quantitative?

Not by default. Ionization efficiency varies across tissue types, so raw intensity is not proportional to concentration in a way that holds across regions. Quantification requires a calibration built into the experiment — typically stable-isotope-labelled standards sprayed across the section so they experience the same local ion suppression — and even then extraction efficiency differences between endogenous and applied analyte remain a known limitation.

How much does an MSI experiment cost to run?

Costs are institution- and vendor-specific, and most academic users access MSI through a shared core facility rather than buying an instrument. The variables that drive the number are instrument time (which scales with pixel count, so with the inverse square of pixel size), matrix and consumables, and staff time for sample preparation and data analysis. Ask your core for their per-hour or per-section rate and for a realistic acquisition time estimate at your intended pixel size before budgeting.

Can I run MSI and then stain the same section?

Frequently yes for MALDI: the matrix is removed with graded ethanol washes and the section is then stained. This gives the best possible co-registration because no serial-section deformation intervenes. Verify compatibility for your specific tissue and matrix before committing the only section you have.