SpaRank

SpaRank is a spatial-transcriptomics deconvolution foundation model developed by Wei Lan's lab at Guangxi University and posted to bioRxiv in May 2026. Spatial transcriptomics measures gene expression in situ across a tissue, but many platforms capture each spatial "spot" as a mixture of several cells. Deconvolution is the task of estimating the cell-type composition within each spot, typically by referencing a single-cell atlas. SpaRank addresses a persistent limitation of existing deconvolution methods: they generally must be retrained for every new tissue, disease context, or sequencing platform, because the absolute expression values they rely on shift across batches and technologies.

The central innovation is a rank-based encoding of spatial spots. Rather than feeding raw or normalized expression magnitudes, SpaRank represents each spot by the relative ordering of its genes, a representation that is far more stable across the batch effects and platform differences that otherwise break transfer. This stability enables a genuine pretrain-and-transfer paradigm: a single model pretrained on one reference can be applied to distinct downstream contexts without retraining, in contrast to the per-context retraining that prior approaches require.

The preprint demonstrates cross-context transfer in two directions, applying a model pretrained on a lymphoid atlas to diverse tissues and platforms, and a model pretrained on a breast atlas to both normal and malignant tissue. SpaRank also includes a multimodal extension that incorporates additional data through a gated-fusion mechanism.

#Key Features

• Rank-based spot encoding: Represents each spatial spot by the relative ranking of gene expression rather than absolute magnitudes, a transformation that is robust to batch effects and platform differences and underpins cross-context transfer.
• Pretrain-transfer paradigm: A single pretrained model generalizes to new tissues, disease states, and platforms without retraining, departing from the per-context retraining that conventional deconvolution methods require.
• Cross-context generalization: Demonstrated transfer from a lymphoid atlas to diverse tissues and platforms, and from a breast atlas to both normal and malignant tissue.
• Multimodal gated fusion: An optional extension integrates additional modalities through a gated-fusion mechanism, broadening the data the model can leverage.

#Technical Details

SpaRank is framed as a foundation model for spatial-transcriptomics deconvolution, built around its rank-based representation of spatial spots and a pretrain-then-transfer training regime. The preprint reports cross-platform and cross-tissue evaluations using lymphoid- and breast-derived reference atlases as pretraining sources, with the multimodal variant adding a gated-fusion pathway for auxiliary modalities. Detailed architecture specifications and the model's parameter count are not stated in the available preprint, so these are not reported here.

The work is a version 1 bioRxiv preprint (DOI 10.64898/2026.05.09.723936) released under a CC BY-NC license and has not yet undergone peer review. As of this writing, no public code repository, trained weights, or HuggingFace release has been confirmed, which limits independent reproduction and direct hands-on use.

#Applications

SpaRank is aimed at spatial-biology researchers who need cell-type deconvolution across heterogeneous datasets without building and validating a new model for each tissue, disease, or platform. Its transfer paradigm is especially relevant for studies that span multiple tissue types or compare normal and diseased states, such as tumor microenvironment characterization, where a single deconvolution model applied consistently across contexts reduces methodological variability. The multimodal extension further suits settings where complementary data can be fused to sharpen composition estimates.

#Impact

By introducing a rank-based encoding that makes deconvolution transferable, SpaRank reframes a task that has historically been treated as context-specific into one amenable to a shared, pretrained foundation model. If the cross-context results hold up under peer review and independent testing, this could meaningfully lower the per-study cost of spatial deconvolution and improve comparability across datasets. As an unreviewed preprint without a confirmed public code or weights release, its real-world adoption and validation remain to be established.