One H&E slide, three transcriptomes: a multi-scale graph network beats DeepSpot2Cell in cell-level PCC and AUROC
Synopsis
Spatial and single-cell transcriptomics reveal where genes are expressed in tissue, but each slide is expensive and scarce, while routine H&E whole-slide images are abundant. A multi-scale graph neural network trains on TCGA bulk data, Visium spatial data and Xenium single-cell data together, using consistency losses to keep predictions aligned across the three scales, and predicts bulk, spatial and single-cell expression directly from slide images. It performs best at slide level and moderately at patch level, with single-cell prediction the hardest; on Xenium its cell-level PCC and AUROC exceed DeepSpot2Cell, while DeepSpot2Cell leads at patch level.
Interpretation
A single model can now produce a slide-wide average expression, spot-level spatial expression and single-cell expression from one H&E whole-slide image, instead of training three separate models. Earlier deep learning models inferred only partial gene expression from pathology slides, with limited gene coverage, weak generalization across tissues and low accuracy in reconstructing the full transcriptome; single-cell approaches typically treat a spot as a bag of cells and learn cell expression indirectly from spot labels through multiple instance learning. On Visium the overall bulk PCC is strong while patch-level PCC drops to a moderate level, a gap consistent with the inherent difficulty of predicting spatial transcriptomics; on Xenium both slide-level and patch-level prediction stay strong, and cell-level PCC is smaller but still a moderate positive correlation.
The model treats tissue as nested graphs of cells, patches and slides, passing information upward level by level and constraining predictions at the three scales to agree. Visual features come from the pathology foundation model UNI2-h, pretrained only on general histopathology images and never exposed to spatial transcriptomics data; cell segmentation and classification use Classpose, with five classes: neoplastic, connective, inflammatory, dead and epithelial. Training combines TCGA, Visium and Xenium samples across lung, kidney and breast tissue, with five-fold cross-validation split at slide level so test slides and patients never appear in training.
Removing the cross-scale consistency losses makes cell-to-patch agreement on Visium collapse from a high level, showing that this constraint is what keeps the three scales aligned. Removing the feature reconstruction objective lowers Visium patch-level AUROC and the fraction of significant genes; removing TCGA samples instead raises bulk PCC on Visium and Xenium but degrades feature reconstruction quality. These ablations compare the full model with four variants on the same Visium, Xenium and TCGA cohorts, using PCC, AUROC, feature cosine similarity and the fraction of genes significant at FDR below 0.05.
Perspective
The demonstrated scope is TCGA, Visium and Xenium samples from lung, kidney and breast tissue, evaluated with five-fold cross-validation split at slide level; single-cell predictions can only be directly evaluated for genes actually measured by the Xenium panel, and TCGA spatial predictions have no ground-truth spatial transcriptomics for comparison.
Single-cell prediction on Xenium is only moderately correlated, and sparse cell counts make the cell-level graph visibly noisier; Visium carries strong slide-level batch effects, and the small breast cohort limits statistical power; the models differ in available supervision, which limits benchmark conclusions; adding TCGA to training improved feature reconstruction but lowered bulk PCC on the spatially resolved datasets.