Spatial Proteomics

Imaging dozens of protein markers on one tissue section at single-cell resolution — which breaks every model built for H&E, because the input stops being three fixed colour channels and becomes N arbitrary named ones.

Imaging dozens of protein markers on one tissue section at single-cell resolution — which breaks every model built for H&E, because the input stops being three fixed colour channels and becomes N arbitrary named ones.

What it is

An IHC slide shows one marker. Spatial proteomics shows twenty to sixty on the same section, registered to the same cells, at single-cell resolution. Instead of “is this cell Ki-67 positive”, the question becomes “what is this cell’s whole protein profile, and what is it sitting next to”.

The platforms differ in how they get there — CODEX and PhenoCycler cycle fluorescent antibodies on and off the same section; MIBI and IMC use metal-tagged antibodies read by mass spectrometry; CyCIF cycles immunofluorescence — but the output is the same shape: a stack of single-channel images, one per marker, plus a nuclear channel (usually DAPI) for segmentation.

This is not the same as Cross-Stain Registration, which aligns separate sections stained differently. Here there is one physical section and every marker is already in register, which is precisely what makes cell-level co-expression answerable.

Three questions follow naturally, and they map onto the three scales the field works at:

• Cell — what type is each cell? (phenotyping, from its marker profile)
• Neighbourhood — which cell types recur together, and where? (spatial niches, usually found by unsupervised clustering rather than defined in advance)
• Patient — do those niches predict outcome or treatment response?

The middle one is the reason the modality exists. Whether a cytotoxic T cell is inside a tumour nest or stuck at its edge is invisible to a bulk assay and invisible to any single stain, and it is the kind of thing that turns out to matter.

Why an H&E foundation model cannot simply be pointed at it

Foundation Models in Pathology describes encoders trained on enormous numbers of H&E tiles. The input to those is fixed: three colour channels, always in the same order, always meaning the same thing. A spatial-proteomics input is not like that in three ways.

The channel count varies. One panel has 18 markers, another 57. A network with a fixed first-layer channel count cannot take both.

The channels are named, and the names carry the meaning. Channel 4 is not “the fourth colour”, it is CD8 — and in the next study channel 4 is panCK. Order is arbitrary; identity is everything. A model that ignores which marker is which has thrown away most of the signal.

Panels barely overlap between studies. Two labs studying the same disease may share only a handful of markers. A model that can only run on the exact panel it was trained on is useless outside the lab that made it.

The word for a model that survives all three is panel-agnostic: it takes an arbitrary set of named markers and produces a representation anyway. That is the central architectural problem of the field, and it has no counterpart in H&E work.

The consequence is practical and slightly unglamorous: the marker vocabulary becomes part of the model interface. Channel names arrive as whatever the operator typed — CD3e, CD3, CD3-e, cd3_epsilon — and something has to map those onto canonical identities before any model sees them. CORAL treats this as a first-class stage rather than a preprocessing afterthought: an editable marker map that is the single source of truth, re-validated at the start of every pipeline stage, with unrecognised names flagged for review instead of silently dropped, against a canonical registry of 287 markers annotated by compartment and family. That design is worth knowing about even for people who will never run it, because the failure it prevents — a panel quietly mis-mapped, invalidating everything downstream — is not one that announces itself.

What the models are, and what is actually established

KRONOS is the published example: self-supervised training on over 47 million patches spanning 175 protein markers, 16 tissue types and 8 fluorescence platforms, evaluated across 11 independent cohorts (arXiv:2506.03373, June 2025 — still a preprint 13 months later). Its notable claim beyond accuracy is segmentation-free patch-level processing — working on tiles rather than requiring every cell to be segmented first, which is what makes cohort-scale analysis tractable and cross-institution comparison possible at all.

Its code answers the panel-agnostic question concretely, and the answer is smaller than the problem sounds. KRONOS is a DINOv2 vision transformer with a marker embedding bolted on: a 512-slot table (num_markers=512, get_1d_sincos_marker_embed) so that a channel is identified by which marker it is rather than by where it sits in the stack. Order stops mattering, arbitrary panels become representable, and the marker vocabulary is now literally a lookup table inside the model. That is worth knowing because it sets the boundary of the claim: a panel-agnostic model is agnostic across the markers it has seen, and what happens to a marker absent from pretraining is not visible in the released inference code. CORAL lists “integrating unseen markers” as a tutorial that does not exist yet, which is suggestive. [unverified]

KRONOS2 was announced in July 2026 as its successor, with a toolkit, CORAL. At the time of writing it has no technical report, and the citation offered points at the KRONOS1 preprint — so its training data, architecture and marker vocabulary are undocumented. Both models are gated on Hugging Face and licensed CC BY-NC-ND. [unverified]

One result from CORAL’s own benchmark deserves to be carried around, because it is the sort of thing that gets lost between a table and a headline. Across three cell-phenotyping tasks, simply averaging the marker intensities in a patch scored within two balanced-accuracy points of KRONOS2 (0.702/0.737/0.790 against 0.720/0.742/0.796). On the survival task the gap was large (C-index 0.694 against 0.535). Read together: on that evidence the foundation model earns its keep at the tissue and patient level, and at the cell level a naive intensity summary is nearly as good. If someone is choosing a method for cell phenotyping alone, the baseline is the thing to beat and it is not far behind. Caveats on both halves — self-benchmark, and n = 80 for the survival column — are on CORAL.

Why it matters for my work

The department does not do this, and that is the point of writing it down. Nothing in sources/projects/ or wiki/datasets/ records a CODEX, PhenoCycler, MIBI or IMC platform. So this page is not a how-to; it is a map of what the entry cost would be, written while a concrete example was in front of me.

The honest assessment of transferability, in both directions:

What transfers from here. Marker interpretation — knowing what CD8, panCK or Ki-67 mean in tissue, and what a plausible profile looks like — is exactly the expertise a pathologist has and a computational group usually lacks. Panel design is a pathology question before it is a technical one. Interobserver Agreement applies unchanged: cell phenotyping needs a reference standard, and someone has to make it.

What does not. The imaging is not scanner work. Cycling chemistry, autofluorescence, channel-to-channel registration and batch effects across cycles are a different operational world from Whole Slide Imaging, and none of the group’s Scanner Log Analysis or WSI Quality Control machinery reaches them.

Two connections are more immediate than the modality itself.

Everything here is Intratumoral Heterogeneity made measurable. That page records marker expression varying within one tumour so the answer depends which piece was sampled — for HER2 Intratumoral Heterogeneity that variation is the object of study and the tool is single-marker IHC on serial sections. Spatial proteomics is the instrument that measures that variation directly, with co-expression preserved per cell. If that project ever needs to ask what else is happening in the heterogeneous regions, this is the technique that answers it.

The niche-discovery pattern is method-agnostic. Unsupervised clustering of local composition into recurring neighbourhoods is a way of thinking about tissue, not a piece of equipment. The same question — which arrangements recur, and do they predict anything — is askable of any spatial data the group already has, including annotated H&E.

How it connects

CORAL — the worked example this page was written from: the pipeline, its marker-map design, and its licence problem.

KRONOS — the published model underneath all of it, where the panel-agnostic claim is a 512-slot marker embedding, and the one artefact in this family with a paper and real numbers behind it.

Foundation Models in Pathology — the same frozen-encoder idea one modality across; the differences are in the input, and they are larger than they look.

Cross-Stain Registration — the alternative when you have several stained sections rather than one multiplexed section, and the reason it can never match cells the way this does.

Intratumoral Heterogeneity — the biology this measures directly, and the group’s existing project that most obviously wants it.

Ki-67 Proliferation Index — the single-marker version of the same question, and a reminder that scoring reproducibility is hard even with one marker, let alone fifty.

QuPath Annotation Workflow — the practical bridge: multiplex tooling exports GeoJSON masks for hand correction in QuPath, so the annotation habits here already apply.

Patch Extraction — the same tiling decisions, with an extra axis. Patch size and µm/pixel still set the ceiling, and now the marker subset does too.

Concordance Index — the metric the patient-level claims in this area are reported with, and the one that is least stable at the cohort sizes involved.

Open questions

• Is there any prospect of multiplex data here — a collaborator, a core facility, a platform in the building? Unrecorded. [unverified] It is the question that decides whether any of this is worth revisiting.
• What does an entry-level panel actually cost per slide, in reagents and instrument time? Not investigated, and it is the number that would make the conversation concrete.
• Would HER2 Intratumoral Heterogeneity answer a question it cannot currently answer if a handful of its cases were run multiplexed? That is the sharpest version of “is this worth it” for this group, and it is answerable by asking rather than by buying anything.
• Does the segmentation-free patch-level approach have a counterpart worth borrowing for H&E cell-level work here — i.e. skipping Cell Detection and Counting for questions that do not strictly need per-cell output? Worth thinking about; nothing in this wiki has asked it.