Spatium

Spatial proteomics measures protein abundance in single cells while preserving their position in intact tissue, but every experiment does so through a small, hand-picked antibody panel. Two datasets of the same tissue may share only a handful of markers, and different platforms report expression on different intensity scales. The result is a fragmented measurement space, where analyses are built per dataset with task-specific models that cannot transfer to an unfamiliar panel, and where learning reusable representations has lagged behind the equivalent effort in transcriptomics.

Spatium, developed at UTHealth Houston, is a foundation model built for exactly this setting. It is trained on more than 51 million cells drawn from multiple spatial proteomics platforms, and it learns the intrinsic co-expression hierarchies among proteins rather than a mapping from one fixed panel to one fixed output. Those hierarchies are what make the representation robust to panel composition and to measurement scale, so cells profiled with different antibody sets land in a shared, comparable embedding space.

The paper describes Spatium as a protein language foundation model, and the analogy is to models like scGPT rather than to amino-acid sequence models: a cell is presented as a set of measured markers with associated expression levels, and the model learns over those expression tokens. That distinguishes Spatium from image-based spatial proteomics models such as KRONOS, which encodes multiplexed image patches, and from panel-agnostic phenotyping models such as DeepCell Types, which reads per-cell image crops. Spatium instead operates on the quantified single-cell expression table that segmentation produces — the form in which most spatial proteomics data is archived and shared.

#Key Features

• Panel-robust representations: The learned embedding captures cell identity in a way that holds up across different panel compositions and measurement scales, making cells from separately designed experiments directly comparable.
• 51-million-cell multi-platform corpus: Pretraining spans more than 51 million cells collected across multiple spatial proteomics platforms, giving the model exposure to the marker overlap patterns and intensity distributions that vary between technologies.
• Co-expression hierarchies as the learning signal: Rather than fitting panel-specific decision boundaries, Spatium models how proteins co-vary, which is the structure that generalizes when the available markers change.
• One backbone, several tasks: Cell type recovery, spatial microenvironment characterization, and reconstruction of unmeasured proteins are all served by the same pretrained representation with lightweight task-specific adaptation.
• Interpretable marker signatures: The reported cell identities carry marker expression patterns consistent with known biology, and the recovered microenvironments are characterized by coherent marker enrichment rather than opaque cluster labels.

#Technical Details

Spatium is a transformer trained over expression tokens: each cell's measured proteins and their expression levels are encoded as input, so which markers a panel happens to include becomes part of the model's input rather than a fixed input dimension. Pretraining covers more than 51 million cells across multiple platforms, and the resulting representation is reused across downstream tasks through lightweight task-specific adaptation instead of a separate model fit per dataset — the property that makes it a foundation model rather than a collection of per-panel classifiers.

The reported evaluations exercise three capabilities. Spatium recovers cell identities whose marker expression patterns match established biology; it resolves functionally distinct spatial microenvironments with coherent marker enrichment signatures; and it reconstructs missing protein measurements while preserving biologically meaningful expression structure. Performance is described as stable and interpretable across these settings. Project documentation puts the pretraining corpus at 70 human studies spanning CODEX, IMC, CyCIF, and MIBI across 18 tissues and 18 disease conditions.

Code is released under Apache-2.0 and includes the transformer implementation, the masked-token pretraining loop, and fine-tuning heads; a pretrained checkpoint is committed to the repository. The accompanying documentation site is still under construction, and Spatium is currently a preprint that has not yet been peer reviewed.

#Applications

Spatium is aimed at groups that already generate or aggregate spatial proteomics data and want to analyze it in bulk rather than one cohort at a time. Cell typing without per-dataset gating or retraining is the immediate use, particularly for archives assembled from studies that used incompatible panels. Microenvironment characterization supports tumor immunology and tissue atlasing work, where the question is which cellular neighborhoods recur and what marker programs define them. The imputation capability is useful when a marker of interest was absent from an existing panel, offering a computational estimate in place of re-staining a precious specimen — an inference to be treated as a hypothesis, not a measurement.

#Impact

Spatium extends the foundation-model approach that reshaped single-cell transcriptomics to the protein layer of spatial biology, and it does so by treating panel heterogeneity as the central modeling problem rather than a preprocessing nuisance. That framing matters because panel fragmentation, more than data volume, is what has kept spatial proteomics analyses siloed. By showing that one representation can serve cell typing, niche discovery, and imputation with only lightweight adaptation, the work argues that the expression-table view of spatial proteomics is rich enough to support a shared model — a complement to the image-based encoders that dominate the field. The caveats are those of any new preprint: the evaluations are computational, imputed markers await experimental confirmation, and independent benchmarking against per-dataset methods has yet to accumulate.