scGPT-spatial

Single-cell foundation models learn from dissociated cells, which means they learn nothing about where those cells sat in the tissue. Spatial transcriptomics recovers that context, but it fragments into incompatible protocols: Visium spots pool RNA from several neighbouring cells across the whole transcriptome, while imaging platforms like MERFISH and Xenium resolve single cells but only for a few hundred targeted genes. A model pretrained on scRNA-seq has no way to reconcile those distributions, and no notion of what "next to" means.

scGPT-spatial, from the Bo Wang lab at the University Health Network and the Vector Institute, with collaborators at the Arc Institute and UCSF, is a specialized descendant of scGPT produced by continual pretraining on spatial data rather than training from scratch. It keeps scGPT's learned gene and cell representations and adds three things the parent model lacks: a mixture-of-experts decoder that routes each sample to protocol-specific experts, a spatially-aware sampling strategy that draws training batches from contiguous tissue neighbourhoods, and a neighbourhood-based reconstruction objective that forces the model to predict a spot's expression from its surroundings.

The team also curated SpatialHuman30M, the corpus that made this possible: 30 million spatial transcriptomic profiles spanning four assay types.

#Key Features

• Protocol-aware decoding: The mixture-of-experts decoder sends spots from each sequencing protocol to specialized experts, so Visium and Xenium distributions are harmonized rather than averaged together.
• Spatial neighbourhoods as training units: Coordinate-based sampling draws contiguous regions from individual slides instead of random cells, which also enables spatially-masked reconstruction.
• Co-localization without coordinate inputs: The neighbourhood objective teaches spatial structure implicitly, so embeddings generalize across slides that never share a coordinate frame.
• Cell-type deconvolution: Feeding its spot embeddings into a Tangram-style factorization pipeline lifts dominant cell-type Macro F1 to 0.58 across ten Visium Human Breast slides.
• Contextualized gene imputation: Missing genes in targeted imaging panels are imputed from a Visium reference, recovering expression patterns that retrieval-based methods blur.

#Technical Details

SpatialHuman30M comprises roughly 30 million cells and spots from 821 unique spatial slides covering more than 20 organs and tissues in healthy, cancerous, and diseased contexts, drawn from four assays: Visium, Visium HD, MERFISH, and Xenium. Continual pretraining starts from the released scGPT checkpoint and adds MoE decoders on top of the shared transformer encoder. In multi-slide and multi-modal integration, evaluated zero-shot against PCA and Seurat v4, scGPT-spatial reaches an AvgBIO score of 0.86 and a 10–15% improvement over benchmarks on cell-type annotation structure. Fine-tuned on individual slides it outperforms graph-based spatial domain methods including SpaGCN and stLearn, both of which additionally consume imaging features. For deconvolution across ten Visium Human Breast slides it averages 0.58 Macro F1 on dominant cell types, a 6% margin over Tangram, and for MERFISH imputation against an scRNA-seq reference it holds a median Pearson correlation above 0.6 across six spatially differentiated genes. Code is MIT-licensed; V1 weights and the SpatialHuman30M source table are hosted on figshare, and the preprint itself is CC-BY-NC-ND.

#Applications

The immediate value is for labs assembling spatial atlases from heterogeneous sources — different platforms, different slides, different studies — where batch and protocol effects otherwise dominate the embedding. Deconvolution matters for Visium experiments, where each spot is a mixture and cell-type proportions are the quantity of interest. Imputation matters in the other direction, letting an imaging experiment restricted to a few hundred probes borrow whole-transcriptome context from a sequencing reference.

#Impact

scGPT-spatial demonstrates that continual pretraining is a practical route to specialization: rather than paying for a spatial foundation model from scratch, it reuses a single-cell model's gene representations and spends its compute on what is genuinely new. The mixture-of-experts decoder is the transferable idea here, giving one backbone a principled way to serve assays with very different noise and sparsity characteristics. Limitations are worth noting: the corpus is human-only, the model inherits whatever biases scGPT's scRNA-seq pretraining carried, the preprint has not been peer reviewed, and its CC-BY-NC-ND terms are narrower than the MIT license on the code.