STRESS

Most published spatial transcriptomics data was collected at a resolution coarser than the biology it was meant to describe. A 10x Visium spot spans 55 micrometers and pools several cells; the gaps between spots are measured, in aggregate, not at all. Computational resolution enhancement tries to fill in that missing signal, but the methods that do it best generally require something extra — a paired H&E image, a matched single-cell reference, or both — which many archived datasets simply do not have.

STRESS, from the Institute of Biopharmaceutical and Health Engineering at Tsinghua University's Shenzhen International Graduate School, takes the constrained version of the problem: enhance resolution using the spatially resolved transcriptomics data and nothing else. It reframes a capture-spot array as a three-dimensional tensor of genes by spatial rows by spatial columns, and treats enhancement as a dense prediction task over that tensor, emitting expression profiles for the positions between the original spots rather than subdividing the spots themselves. The sequence backbone is a state space model in the Mamba family, scanned along three axes — gene, height, and width — so that long-range dependencies across thousands of genes and hundreds of spatial positions are captured at linear rather than quadratic cost.

The design pays off as transferability. A model trained on dorsolateral prefrontal cortex Visium data was applied unchanged to independent Visium HD and Slide-seq V2 datasets, with fine-tuning available as an option rather than a requirement. The preprint was posted in June 2025.

#Key Features

• No auxiliary modality required: Enhancement depends only on the SRT matrix and spot coordinates, so it applies to archived datasets that lack paired histology or matched single-cell references.
• Tri-orientated Mamba blocks: TSMamba scans the tensor along the gene, height, and width directions, letting one block model gene-gene co-expression and spatial continuity within a single linear-time pass.
• Predicts gaps, not subdivisions: Output is expression at the unmeasured positions surrounding each spot, quadrupling the number of positions rather than splitting an existing measurement into pieces.
• Self-supervised training signal: Down-sampled versions of real arrays serve as input with the original resolution as ground truth, so no separately generated high-resolution reference is needed.
• Cross-platform generalization: Parameters learned on Visium data transfer to Visium HD and Slide-seq V2 without retraining, with optional fine-tuning for further gains.

#Technical Details

The architecture is an encoder-decoder: a stem projection layer feeds four stacked TSMamba blocks with hierarchical down-sampling, and a U-Net-style decoder with symmetric up-sampling and residual blocks recombines multi-level features. Two auxiliary components supplement the state space scan — a Spatial Auxiliary Convolution module for local spatial context and a Gene-constrained Dilatation block for spot-level detail. An input tensor of G genes by H by W becomes G by 2H by 2W, a four-fold increase in spatial positions.

Training and evaluation used DLPFC 10x Visium, HER2ST breast cancer, and Stereo-seq mouse organogenesis data with leave-one-out cross-validation. On DLPFC, STRESS improves peak signal-to-noise ratio by 225.7% and structural similarity by 8.0% over nearest-neighbor interpolation while reducing RMSE by 8.7%, and beats BayesSpace on the same targets. Applied without retraining, it reaches a mean PSNR of 41.01 and SSIM of 0.9653 across six Visium HD samples, and a mean Pearson correlation of 0.5262 on Slide-seq V2 mouse hippocampus data covering 51,367 spots. Downstream, enhanced resolution sharpens spatial domain segmentation: silhouette score on an invasive ductal carcinoma sample rises from 0.061 to 0.512.

#Applications

The practical use is retrospective. Groups sitting on Visium or Stereo-seq datasets collected before higher-resolution platforms existed can raise effective resolution without new tissue or new sequencing, sharpening cortical layer boundaries, tumor margins, and other spatial domains that coarse spots blur. It is complementary to histology-conditioned imputation models such as FOCUS, which reach finer detail but require the paired imaging or single-cell data that STRESS deliberately does without.

#Impact

STRESS extends state space sequence models into spatial omics, where the combination of many genes and many positions makes attention costly, and it establishes that useful resolution enhancement is possible from expression data alone. The authors are explicit about the limits: mRNA diffusion and sparsely expressed genes can produce false-positive imputed signal; compute constraints restricted training input to a highly variable gene subset rather than the full transcriptome; and predictions are extrapolations from spatial and expression structure, so individual gene calls in unmeasured positions should be treated as hypotheses. The model was developed for the 10x Visium geometry, though it can be retrained for other platforms. It remains a preprint, and the released code carries no license file, leaving reuse terms undefined.