Spatial-proteomics studies do not all measure the same proteins. One cohort may contain 40 markers, another 60, and a third may use a different imaging technology altogether. VirTues was built around that mismatch. It learns from the proteins that are present, their sequence-derived identities and the arrangement of cells in tissue, then builds a common representation that can support several downstream tasks.[1]
The core training data covered 15 imaging-mass-cytometry cohorts, 3,102 patients and 146 markers. An extended evaluation reached 32 cohorts, more than 5,100 patients, 239 markers and four spatial technologies. That breadth is the reason the system is described as a tissue foundation model rather than a classifier for one cancer or one staining panel.[1]
One representation, several levels of tissue
| Scale | Task | What the output represents |
|---|---|---|
| Protein marker | Reconstruct a masked or unmeasured marker | An inferred intensity pattern, not a new laboratory stain.Source 1 |
| Cell | Segmentation and cell-type classification | A model-assigned boundary or identity from the available image context.Source 1 |
| Neighbourhood | Identify multicellular niches | A learned description of which cells and markers co-occur locally.Source 1 |
| Patient | Stratification and response analysis | A retrospective association between the tissue representation and an outcome.Source 1 |
Protein language helps bridge missing panels
VirTues does not treat every marker name as an arbitrary label. It embeds the protein sequence, giving the model information about relationships between proteins even when two cohorts do not contain exactly the same panel. Spatial tokens then represent local tissue regions, while factorised attention lets the system reason across markers and locations without treating the whole slide as one flat image.[1]
In marker-reconstruction tests, the paper reports mean correlations of about 0.800 when individual measurements were masked, 0.700 when a marker was hidden and 0.668 when an entire local niche was masked. These are useful reconstruction results, not proof that a virtual stain is interchangeable with the physical assay.[1]
The model can transfer, but transfer is not magic
The authors also tested zero-shot transfer to CODEX and Orion datasets and reported cell-typing performance above several baselines. The model’s advantage is that it has seen many tissues, markers and technologies. The remaining risk is domain shift: preparation protocols, scanner behaviour and patient populations can differ in ways the representation has not learned.
Selected reported results
| Measure | Value | Context |
|---|---|---|
| Core training scale | 3,102 patients; 146 markers | Across 15 imaging-mass-cytometry cohorts.Source 1 |
| Extended evaluation | >5,100 patients; 239 markers | Across 32 cohorts and four technologies.Source 1 |
| ER-positive breast cancer cohort | 541 patients | The retrospective risk groups contained 265 patients with 99 deaths and 276 with 58 deaths over 21 years.Source 1 |
| Pre-treatment TNBC response task | Macro-F1 0.676 | The on-treatment and post-treatment analyses reported different values and use different clinical timepoints.Source 1 |
Why this matters beyond one benchmark
A tissue model becomes more useful when researchers do not have to rebuild it for every marker panel. VirTues suggests a route to that interoperability. The most credible near-term role is research support: harmonising cohorts, prioritising markers, describing niches and generating hypotheses. Turning those capabilities into clinical decisions would require prospective validation, stable assay pipelines and evidence that the inferred representation adds value beyond the measurements already available.
Sources & further reading
- VirTues: a virtual tissue foundation model for spatial proteomics Nature · 5 Aug 2026 · Accessed 16 Oct 2026