Multicellular spatial transcriptomics gets split into single cells: st2sc tracks attention gradients to find cells, then regresses gene counts
Synopsis
Each spot in spatial transcriptomics typically mixes signals from several cells, and earlier methods leaned on cell-type prediction or external single-cell references, leaving absolute per-cell gene counts imprecise. st2sc separates the task into cell identification and gene counting: it traces gradient flow on attention maps from a Vision-Gene Transformer to propose candidate cells, then a set of per-gene regressors predicts each cell's expression counts under weak supervision. On a Visium-like simulation built from Xenium breast cancer data, its per-cell mean squared error is 3.6833, below SpatialScope's 18.2552 and scResolve's 65.3402; on contested 2-micron bins in Visium HD, 18.6% of tested bins gave both cells their own marker counts.
Interpretation
Spatial transcriptomics at multicellular resolution can be decomposed into per-cell gene expression counts, not only cell-type proportions. Earlier SpatialScope depends on a matched external single-cell reference and scResolve depends on method-specific parameters such as metagenes, and both treat relative cell-type prediction accuracy rather than absolute accuracy of inferred gene counts as the main measure. On a Visium-like simulation built from 10x Xenium FFPE human breast data (3,971 spots, 313 genes), st2sc reaches a per-cell mean squared error of 3.6833, against 18.2552 for SpatialScope and 65.3402 for scResolve.
Cell positions are read directly from attention maps of the histology image without a predefined background threshold. Conventional image-based gradient flow tracking first segments the background by thresholding, and the threshold choice introduces performance bias. On the Visium-like simulation, st2sc identifies 49,242 cells, 69.5% of the cells whose centers lie within a spot, versus 39,326 (55.5%) for SpatialScope and 19,687 (27.8%) for scResolve.
Each gene gets its own small regressor, and the per-cell predictions are summed to match the observed spot value, so training works with spot-level labels only. Spot-level expression is the sum of several cells' expression, so no direct supervision exists for an individual cell's counts. The regressors use a Huber loss that is more robust to outliers, and on the Visium-like simulation cell-type accuracy is 69.50% with per-spot cell-type correlation 0.8911, against 52.77% and 0.8584 for SpatialScope and 65.72% and 0.7534 for scResolve.
Perspective
The evaluation covers three datasets: a Visium-like simulation built from Xenium human breast data (3,971 spots, 313 genes), 2-micron expression bins from mouse small intestine Visium HD, and one human IPF lower-lobe Visium sample (2,633 spots, 17,956 genes). Visium HD has no ground-truth count assignment, so the fraction of cell-type marker counts landing in the matching cell serves as a proxy.
Postprocessing of attention gradient flow tracking does not fully remove spurious instances or merge all fragmented instances; the authors suggest further postprocessing using the angle map and note that the model currently only shows potential for capturing cell boundaries rather than precise segmentation.
