Spatial Proteomics: Methods for Tissue Mapping

Most spatial proteomics projects are not decided by the instrument. They are decided by two choices made before any tissue is sectioned: whether you need a targeted or a discovery answer, and how much work you are willing to put into the antibody panel. Plex count — the number people compare platforms on — is rarely the binding constraint. As one methods survey puts it, many platforms support 40 or more markers per acquisition, a number that already exceeds what most biological questions require.

This guide covers the platforms, what they actually differ on, and the two failure modes that quietly ruin otherwise well-run experiments: antibodies that were never validated under the conditions they were used in, and segmentation error that manufactures cell phenotypes that do not exist.

First, disambiguate the term

“Spatial proteomics” names two largely separate fields, and searching for one returns protocols for the other.

Tissue-level spatial proteomics — which protein is in which cell, and which cells sit next to which. This is the multiplexed-imaging and imaging-mass-spectrometry world, and it is what this guide is about.
Subcellular spatial proteomics — which organelle or compartment a protein occupies inside a cell. This is the world of organelle fractionation with correlation profiling (LOPIT and hyperLOPIT), proximity labelling, and the Human Protein Atlas subcellular resource.

Both are legitimately called spatial proteomics. They share almost no methodology. If a paper, a core facility quote, or a reviewer comment seems to be answering a different question than the one you asked, this ambiguity is usually why. There is a short section on the subcellular branch near the end.

The fork that determines everything: targeted or discovery

Every tissue-level method falls on one side of a line, and the line is not negotiable by budget.

Targeted: antibody-based multiplexed imaging

You stain a section with a panel of tagged antibodies and image them. You get single-cell resolution, intact tissue architecture, and the ability to say which cell types are adjacent to which. You see only what you put on the panel. If the biology depends on a protein you did not include, the experiment cannot tell you — and you will not know it failed.

Choose this when you have a hypothesis about specific cell populations or states, and when neighbourhood relationships matter.

Discovery: mass spectrometry with spatial sampling

You either image the tissue directly by mass spectrometry, or you cut regions out and run them as ordinary proteomics samples. You get unbiased coverage — thousands of proteins, no panel to design — at the cost of coarser spatial sampling, more tissue per data point, and much lower throughput.

Choose this when you do not yet know which proteins matter, or when you need quantitative depth on a region rather than phenotype calls on individual cells.

A growing number of studies run both: discovery MS on pooled regions to generate the marker list, then targeted imaging with a panel built from it. If your project can afford two rounds, this ordering wastes far less than starting from a guessed panel.

Antibody-based platforms: what they actually differ on

The differences that matter are detection chemistry, resolution, how much tissue area you can cover, and whether the section survives.

Fluorescence cycling — CyCIF, PhenoCycler (formerly CODEX), Cell DIVE

These build high plex out of ordinary fluorescence by repeating stain–image–remove cycles. Reviews describe t-CyCIF and IBEX reaching more than 50-plex whole-slide, typically via roughly 10–20 cycles to produce 30–60-plex datasets; PhenoCycler is described in the 50–100 marker range using approximately 10–20 hybridisation and imaging cycles with DNA-barcoded antibodies.

The practical advantages are area and accessibility — cyclic fluorescence can cover large tissue areas in a single session, and CyCIF in particular runs on a standard fluorescence microscope with no dedicated instrument. The costs are that every cycle must be registered to every other cycle, autofluorescence must be handled (a real problem in lung and liver), and repeated processing degrades some epitopes while exposing others. One review also notes that quantitative non-linearity in these platforms complicates absolute protein copy-number estimation — treat the output as relative intensity, not molecules per cell.

Metal-tag mass imaging — IMC (Hyperion) and MIBI-TOF

Antibodies carry stable metal isotopes rather than fluorophores. Imaging mass cytometry ablates the section line by line with a laser and reads the released metals by mass cytometry; multiplexed ion beam imaging uses a primary ion beam and time-of-flight detection. Both sidestep autofluorescence entirely, because the readout is mass, not light — a substantial advantage in high-background tissue.

Reported figures vary by source and configuration. IMC is commonly described at approximately 1 µm pixel size with 40-plus simultaneously measured markers. MIBI reaches finer detail: one review reports approximately 100 targets and approximately 260 nm resolution, while a methods guide describes up to 40 markers with resolution down to 260 nm. High-resolution IMC variants have since been published that reach subcellular structures by oversampling with sub-micron step sizes and deconvolution. Take any single number as configuration-dependent and confirm it against the specific instrument and protocol you will be using.

Two constraints are structural rather than incidental. First, both readouts are destructive — ablation and sputtering consume the section, so you cannot go back and re-image the same tissue with a better panel. Second, the point-by-point acquisition is slow relative to wide-field fluorescence, so high resolution is traded against imaged area. Plan the region of interest as a real experimental design decision, not an afterthought at the microscope.

Region profiling with barcode release — GeoMx DSP

This is a different shape of experiment. Antibodies carry UV-photocleavable oligonucleotide barcodes; you choose regions on a stained image, illuminate only those regions, and collect the released barcodes for counting. Vendor documentation describes selection of tissue compartments or cell types using a micromirror array with resolution down to 10 microns, including non-contiguous areas, on FFPE and fresh frozen sections, tissue microarrays and organoids with standard histology staining. The same documentation cites 1,200-plus proteins measurable simultaneously or across serial sections, with an NGS readout supporting up to 40 custom antibodies and an nCounter readout up to 10 custom antibodies.

The trade is explicit: you get high plex and clean quantification per region, but the output is a profile of a region, not of individual cells. It answers “how does this compartment differ from that one” well and “which cells touch which” not at all.

Mass-spectrometry spatial proteomics

MALDI imaging mass spectrometry

The section is coated with matrix and rastered by laser, producing a mass spectrum per pixel. Reviews place MALDI and DESI imaging at roughly 10–50 µm pixel sizes, with whole-section images involving 10⁴ to 10⁶ pixels. For proteins specifically, the tissue is usually digested on-slide and tryptic peptides are imaged: published work has imaged tryptic peptides at 20 µm pixel size in FFPE human breast cancer tissue, at 10 µm with high mass resolution and mass accuracy, and at 5 µm in two-dimensional cell culture.

The honest limitation is identification, not imaging. A MALDI image is a map of mass-to-charge values; turning those into confident protein identifications has historically required a parallel LC-MS/MS experiment on the same tissue, and workflows combining high lateral resolution, accurate mass, efficient digestion and reproducibility in one protocol remain an active problem. Ion suppression also complicates robust quantification. Recent MALDI MS/MS imaging using ion-mobility PASEF is aimed directly at on-tissue identification, but this is a moving field — check the current state before designing around it.

Microdissection plus LC-MS/MS, including Deep Visual Proteomics

The most quantitatively powerful route is still to cut the region out and run it as a normal proteomics sample. Laser capture microdissection excises histologically defined regions, which then go through standard digestion and LC-MS/MS sample preparation and acquisition. Reviews put immunohistochemistry-guided LCM with data-independent acquisition at roughly 10² to 10⁴ cells per region and at most tens to a few hundred regions per day.

Deep Visual Proteomics is the automated form of this idea. In the original Nature Biotechnology report, high-resolution imaging feeds AI-driven segmentation and machine-learning classification of cell phenotypes; the classified cells are then excised automatically by laser microdissection and analysed by mass spectrometry. The published implementation used Leica LMD6 and LMD7 microscopes — the LMD7 excising 1,250 high-resolution contours per hour, equivalent to 50 to 100 cells per sample — with a timsTOF Pro and diaPASEF. Reported depth was more than 5,000 quantified proteins in fallopian tube ciliated cells and 3,653 protein groups across U2OS nuclei classes, from inputs of 80–100 single cells or 250–300 nuclei per phenotype.

At the extreme low-input end, nanoPOTS-style workflows are described at roughly 10–30 cells per day with 30–120 minute gradients. That throughput number is the reason these methods answer focused questions rather than survey questions.

The panel is the experiment

For any antibody-based method, most of the outcome is fixed before the section reaches the platform. The single most common and most expensive mistake is validating antibodies on the wrong material.

Validate every clone in the tissue type and fixation chemistry you will actually use — not on reference cell lines, and not on control tissue from a different organ. Antibodies routinely behave differently under multiplexed conditions than they do in singleplex, so validation carried out after the panel is built is not validation. The general controls logic from conventional immunohistochemistry and immunofluorescence still applies; multiplexing adds requirements on top of it, not instead of it.

Practical steps that published panel-design protocols converge on:

Titrate each antibody singly before assembling the panel, and keep the single-stain images as a reference library — they are what you need later for spectral unmixing and for judging whether a channel is behaving.
Match target abundance to channel sensitivity. Put high-abundance targets on lower-sensitivity channels and low-abundance targets on high-sensitivity ones, so nothing saturates and nothing disappears.
For cyclic methods, treat cycle order as a variable. Heat-induced epitope retrieval degrades some epitopes across cycles and unmasks others; both change signal intensity. Run denaturing controls to work out where in the sequence each antibody belongs.
Run drop-out controls to confirm stripping or removal efficiency between cycles. Residual signal from a previous cycle reads as real expression in the next one.

Do not build the panel from scratch if someone already has

The Human BioMolecular Atlas Program publishes Organ Mapping Antibody Panels (OMAPs) — curated antibody sets validated to work together on a single tissue section for a named organ and a named imaging platform, including tissue preservation details and antibody cycling order. The panels released at Human Reference Atlas v1.2 covered lymph node (IBEX), intestine, kidney and pancreas (CODEX), skin and lung (Cell DIVE), and liver (SIMS); check the current release, since the set has been growing. OMAPs are linked to the ASCT+B tables, so the biomarkers map onto a shared cell-type vocabulary rather than a local one.

Alongside them, HuBMAP publishes Antibody Validation Reports for individual antibodies, recording target and UniProt accession, RRID, host, vendor, catalogue and lot number, positive and negative control tissues and cell lines, isotype controls, a link to the validation protocol on protocols.io, and conjugation and cycling behaviour. The stated validation pillars are specificity (expected spatial pattern, expected cell types and subcellular compartment, colocalisation with orthogonal antibodies, background assessed on controls), sensitivity (signal-to-noise at defined imaging settings and concentration), and reproducibility (consistent across experiments, users and lots).

Even if you deviate from a published OMAP, the AVR fields are a good specification for what your own validation records should contain. Recording the lot number in particular saves projects: lot-to-lot variability is one of the standard explanations for a panel that stopped working.

Where the data goes wrong after acquisition

Clean images do not guarantee clean single-cell data. The dominant analysis failure is segmentation error producing marker signal assigned to the wrong cell.

In densely packed tissue, imperfect cell boundaries let signal from one cell be attributed to its neighbour. The diagnostic signature is characteristic and easy to miss if you are not looking for it: artefactual co-expression of mutually exclusive markers — cells that appear to be simultaneously B cell and T cell, epithelial and immune. If your clustering produces a population like that, suspect lateral spillover before you write it up as a novel cell state.

Compensation methods exist. REDSEA (REinforcement Dynamic Spillover EliminAtion) corrects lateral spillover of cell-surface markers between adjacent cells and has been shown to improve marker recovery and cell-type classification on both ion-beam and cyclic-immunofluorescence images. Its published limitations are worth knowing before you rely on it: it does not perform three-dimensional correction, it cannot resolve cells that physically overlap, it does not address autofluorescence, spectral overlap or isotopic contamination, and its performance depends on segmentation being reasonable in the first place. It is a correction, not a rescue.

Two further issues affect multi-study work. Metal-tag platforms carry isotopic impurity and channel crosstalk that need their own compensation, separate from the spatial spillover problem. And across platforms, variability in resolution, panel composition and staining protocol produces domain-shift effects that limit how well an analysis model trained on one dataset transfers to another — a reason to be cautious about applying a published classifier to your own images without recalibration.

Choosing a method: a short decision path

Do you know which proteins you care about? No → discovery MS (LCM plus LC-MS/MS, or MALDI imaging). Yes → continue.
Do you need single cells, or are regions enough? Regions → barcode-release profiling such as GeoMx DSP, or LCM-based MS. Single cells → continue.
How much area must you cover? Whole slides or many samples → cyclic fluorescence. A defined region where subcellular detail matters → metal-tag mass imaging.
Is your tissue high-autofluorescence? Lung, liver and similar → metal-tag detection removes the problem outright rather than requiring unmixing.
Is the tissue irreplaceable? Ablation-based methods consume the section. Keep serial sections, and do not spend the only block on an unvalidated panel.
Does a published OMAP exist for your organ and platform? If yes, start there. The months it saves in validation usually dominate every other consideration on this list.

The other meaning: subcellular spatial proteomics

If your question is which compartment a protein occupies rather than which cell, the methods are different. The Human Protein Atlas subcellular resource maps protein localisation by immunocytochemistry-immunofluorescence and confocal microscopy in up to three cell lines drawn from a panel of 42; it reports experimental data for 13,603 genes (67% of human protein-coding genes), rising to 17,062 genes (85%) when predictions for secreted and membrane proteins are included, classified across 49 organelles and subcellular structures.

The mass-spectrometry route is correlation profiling — LOPIT and hyperLOPIT fractionate organelles and assign proteins to compartments by their distribution across fractions, giving deep coverage but described as practical for at most tens to low hundreds of samples. Proximity-labelling approaches solve the same problem enzymatically by biotinylating whatever is near a bait protein. None of these preserve tissue architecture, which is exactly why they are a separate field from the imaging methods above.

Frequently asked questions

Does spatial proteomics work on archival FFPE blocks?

Largely yes, and this is one of its strongest arguments. Reviews describe mass-cytometry imaging as robust for archived tissues, GeoMx documentation lists FFPE alongside fresh frozen, and MALDI imaging of tryptic peptides has been performed on FFPE material including long-stored hospital bank samples. The caveat is that fixation chemistry changes epitope behaviour, which is precisely why antibody validation has to be done on FFPE if FFPE is what you will run.

How many markers do I actually need?

Usually fewer than the platform offers. Since most platforms already exceed 40 markers per acquisition — more than most biological questions require — adding targets mainly adds validation burden, cycle count and spillover opportunities. A well-validated 20-marker panel beats a poorly validated 50-marker one every time.

Can I get absolute protein copy numbers per cell?

Not reliably from multiplexed imaging. Quantitative non-linearity in cyclic fluorescence platforms is a documented obstacle to absolute copy-number estimation, and mass-spectrometry imaging faces ion suppression. Treat imaging output as relative intensity that supports comparisons within an experiment, and use microdissection plus LC-MS/MS if quantitative rigour on a region is the point.

Why do my clusters contain cells expressing markers that should be mutually exclusive?

Almost always segmentation error and lateral spillover from neighbouring cells, not a real hybrid population. Check the raw images at those cell boundaries before anything else, then consider a spillover compensation method such as REDSEA — while remembering it cannot correct overlapping cells, autofluorescence, spectral overlap or isotopic contamination.

Should I run spatial transcriptomics instead?

They answer different questions and transcript abundance is an imperfect proxy for protein. Spatial proteomics measures the functional molecule directly, including post-translational states if the panel targets them, but is limited to what you can raise a validated antibody against. Many groups run both on serial sections; several platforms support protein and RNA readouts from the same or adjacent sections.

How much does the panel design phase actually take?

Vendor-independent guidance is consistent that this is the most technically demanding step of the project, involving per-clone titration, tissue- and fixation-specific validation, cycle-order and denaturing controls, and drop-out controls. Budget it as a distinct project phase with its own timeline, not as setup. Starting from a published OMAP for your organ and platform is the main way to compress it.