VirTues represents spatial proteomics at four scales from one backbone: AUROC 0.817 on triple-negative breast cancer response

VirTues (Virtual Tissues), published in Nature in 2026 by Johann Wenckstern and colleagues at the Bunne Lab at EPFL, is a foundation model for spatial proteomics that represents proteins, cells, niches, and tissues from multiplexed marker images through a single pretrained backbone. The published abstract states that one backbone supports marker reconstruction, cell segmentation and typing, niche annotation, spatial biomarker discovery, and patient stratification, including zero-shot annotation across heterogeneous marker panels and datasets. In clinical validation, VirTues-derived biomarkers predict response to anti-PD-L1 chemo-immunotherapy in triple-negative breast cancer and stratify disease-free survival in an independent cohort, outperforming state-of-the-art biomarkers derived from the same datasets as well as current clinical stratification schemes. ASAP works from the Nature paper (DOI 10.1038/s41586-026-10884-y), the authors' preprint arXiv 2501.06039v2, and the official repository bunnelab/virtues to lay out what this model actually unifies.

Tokenizing the marker axis, not just the image

VirTues departs from earlier pathology foundation models in which axis of the data it treats as a first-class citizen. The Nature abstract describes innovations in transformer architecture design, including a tokenization scheme that captures both the spatial and the marker dimensions, and attention mechanisms that scale to high-dimensional multiplex data while preserving interpretability. The pretraining objective is masked autoencoding.

The reason this design is necessary lies in the structure of spatial proteomics data itself. A single image carries expression values for dozens of proteins at every pixel, and which proteins were imaged differs by laboratory and research aim. Treat channels as a fixed-order tensor the way a standard vision model would, and a dataset with a different marker panel simply cannot enter the same model. Because VirTues handles a marker as a token with an identity rather than a positionally fixed channel, panels of different composition pass through the same backbone.

Pretraining scale as reported in the preprint covers 15 imaging mass cytometry (IMC) datasets spanning 8 organ sites and 147 distinct markers. Aggregated, the corpus holds 3,102 patients and 8,887 tissues, more than 259,000 image crops of 256 by 256 pixels, and over 14.5 million segmented cells across nine of those datasets. The checkpoints published in the official repository go further, splitting into a 32-dataset model (virtues-sp32), a 31-dataset model (virtues-sp31), and a 14-dataset IMC model (virtues-imc14).

Cell-type classification gains 6.31% in F1 over KRONOS

The cell classification numbers reported in the preprint earn their weight because they come from a general-purpose backbone rather than a task-specific method. VirTues improved average F1 by 6.31% over KRONOS and by 65.79% over CA-MAE. The tenfold gap between those two baselines reads most naturally as a consequence of CA-MAE never having been designed to handle a marker axis at all.

On cross-dataset evaluation, where the model moves to data it was not trained on, the margins widen. On the Danenberg et al. dataset, B cells improved by 27.9%, myeloid cells by 35.2%, NK cells by 95.6%, and T cells by 30.4%. Double-digit gains across the immune compartment amount to evidence that the model has not overfit to the acquisition conditions of any single dataset.

Tissue-level prediction is reported as macro-F1. Lung cancer subtyping on the Cords et al. dataset reaches 0.856, breast cancer estrogen receptor status on the Danenberg et al. dataset reaches 0.806, and pre-treatment triple-negative breast cancer response on the NeoTRIP cohort reaches 0.676. Placed side by side, the three numbers expose an ordering of difficulty. Reading a label the tissue already carries lands in the 0.8 range, while predicting a treatment response that has not yet happened stays below 0.7.

138 patients in NeoTRIP and 541 in METABRIC

Clinical validation in the VirTues paper splits into two distinct questions, treatment response and survival stratification, evaluated on separate breast cancer cohorts. The first is treatment response. Of 138 breast cancer patients in the NeoTRIP cohort, 67 achieved a pathological complete response, and biomarkers derived from VirTues reached an AUROC of 0.817 against 0.782 for the biomarkers Wang et al. derived from the same data.

The second is survival stratification. Splitting 541 estrogen-receptor-positive patients in the METABRIC cohort into high-risk and low-risk groups produced 99 death events in the high-risk group against 58 in the low-risk group over a 21-year horizon. In an independent cohort, 3 events occurred among 33 low-risk patients and 21 among 45 high-risk patients, with a concordance index of 0.628.

The two results are different in kind. The first is a head-to-head win against the prior best on the same data; the second is evidence of transfer, showing that the risk separation survives a move to a cohort that played no part in fitting. What the Nature abstract emphasizes is the second one, the stratification in an independent cohort.

The axis of pathology foundation models is shifting from morphology to molecules

Competition among pathology foundation models over the past few years concentrated on learning morphology at scale from H&E-stained slides, which are easy to collect in the millions and comparatively standardized. The data VirTues targets has the opposite properties. Multiplexed imaging data is scarce in absolute terms, poorly standardized, and different laboratories image different proteins.

That is where the definition of a foundation model gets tested. A model built because data is abundant and a model built because data is fragmented do not succeed on the same terms. In the first case scale is performance; in the second, the design that binds fragments into one representation space is performance. This is why zero-shot annotation appears in the paper as a central claim rather than a side feature. If a new dataset with a different marker panel can be handled without additional labels, the field's binding constraint loosens.

Seen this way, the architecture matters more than any single row of the results table. A morphology-based model learns what a cell looks like; VirTues learns what a cell expresses. Expecting the latter to have an edge on tasks governed by molecular state, such as treatment response, is reasonable, and the NeoTRIP result does not contradict it. The two approaches are not mutually exclusive, though. A model that uses morphology and molecules together is the natural next step, and this paper is best read as having erected one of the two pillars.

How to read a concordance index of 0.628 against a clinical bar

The number in this paper that deserves the most caution is the independent-cohort concordance index of 0.628. Random prediction sits at 0.5, so 0.628 establishes that a signal exists, but it falls short of what a standalone clinical decision tool requires. Treating it as equivalent in kind to the double-digit gains in cell-type classification, reported in the same paper, would be a category error.

The same applies to AUROC 0.817 against a prior 0.782. The gap is 0.035 on a cohort of 138 patients. When the sample size sits in the low hundreds, how robust a gap of that size is cannot be settled without confidence intervals. The paper is accurate in stating that its biomarkers outperform state-of-the-art alternatives, but a practitioner who translates that into a clinically deployable margin has over-read it.

The genuinely impressive point sits elsewhere: every one of these numbers derives from a single pretrained backbone. This is not the result of training a dedicated cell classifier and a dedicated response predictor and tuning each to its own optimum. Pulling multiple tasks out of one representation and still competing with task-specific methods is the actual achievement here. Reading the result as performance-per-task-count, rather than comparing each figure to a per-task record, fits its character better.

The MIT and CC BY-NC split across checkpoints decides who can deploy

For a pathology AI company or a hospital research group evaluating this model, the first thing to check is not the results table but the license attached to each checkpoint. The official repository releases the code under the MIT License while splitting the weights. virtues-sp32, trained on 32 datasets, and virtues-imc14, trained on 14 IMC datasets, carry CC BY-NC 4.0 and are limited to non-commercial use, while only virtues-sp31, trained on 31 datasets, is MIT-licensed and open to commercial deployment.

That distinction becomes a fork in practice immediately. The repository recommends sp32 for academic use, but for anything with commercial exposure, such as a diagnostic-support product or a contract analysis service, the choice is effectively sp31 alone. How much a single dataset of difference costs in performance is not stated, so any commercially motivated adoption should begin with an in-house reproduction run on sp31.

The second practical issue is data access. The repository provides access through spora, a curated collection of more than 31 spatial proteomics datasets in harmonized formats, and supports converting custom data into the same format. Multiplexed imaging data held by institutions typically uses panels that differ from the public sets, which makes zero-shot annotation on your own panel the first real gate. Measuring zero-shot performance before spending on labeling is the sensible order of operations.

The third is the regulatory path. Validation in this paper is retrospective, not a prospective trial. Any adoption aimed at medical device clearance treats these scores as a starting point, with reproduction on local cohorts and a prospective design required separately. Using the model as a research tool while accumulating in-house data is the realistic route.

Open questions between the preprint and the published version

The figures quoted here come from two sources, which is worth stating plainly. The Nature version is access-restricted, so only abstract-level claims were confirmed there, and the detailed results table comes from the same authors' preprint, arXiv 2501.06039v2. The first version of that preprint was submitted on January 10, 2025, and the revision on December 9, 2025.

Evidence of expansion between the two versions is clear. The preprint describes pretraining on 15 IMC datasets, while the official repository ships checkpoints trained on 32 and 31 datasets. If the pretraining corpus doubled, the published results are likely stronger than the preprint's, but the individual numbers in this article should be read as preprint figures. Readers who need the published values should consult the Nature article directly.

The limits are equally clear. Pretraining leans heavily on imaging mass cytometry, so transfer to other multiplexed imaging modalities remains a separate question. Clinical results concentrate on breast cancer, specifically the triple-negative subtype and an ER-positive cohort, leaving open whether the same margins reproduce in other tumor types. The conditions under which zero-shot annotation holds, meaning how much a new panel must overlap with the 147 markers seen in training, is a threshold each team has to establish on its own data.

Source: ASAP analysis based on the Nature paper The Virtual Tissues foundation model resolves spatial proteomics across scales (DOI 10.1038/s41586-026-10884-y, 2026), the authors' preprint arXiv 2501.06039v2, and the official repository github.com/bunnelab/virtues.