STITCH reconstructs missing spatial transcriptomics data from sparse 3D tissue sections or damaged 2D slices.
Unlike methods that require external reference atlases or matched histological image priors, STITCH learns intrinsic spatial-transcriptomic patterns directly from the observed tissue sample. It uses a decoupled generative design that separates spatial morphology restoration from transcriptomic generation, enabling scalable reconstruction across diverse spatial transcriptomics platforms.
- Multidimensional reconstruction for both 3D cross-slice gaps and 2D in-slice tissue damage.
- Internal learning paradigm without external atlas priors or matched histological images.
- Decoupled architecture separating spatial coordinate reconstruction and gene expression generation.
- Point-wise Gene Flow with linear computational complexity for scalable transcriptomic generation.
- Cross-platform compatibility across single-cell and spot-level spatial transcriptomics datasets.
- Million-cell-scale atlas reconstruction on a single commodity GPU.
STITCH consists of three core modules:
| Module | Role |
|---|---|
| Encoder-Decoder | Compresses high-dimensional gene expression into a topology-preserving latent space |
| Structure Flow | Reconstructs missing spatial coordinates for 3D gaps and 2D damaged regions |
| Gene Flow | Generates transcriptomic profiles at reconstructed spatial locations |
The default STITCH framework uses a point-wise Gene Flow design for scalability. We also provide a neighborhood-enhanced variant, STITCH-n, to evaluate the effect of local spatial neighborhoods on generative fidelity and computational cost.
STITCH supports:
- 3D virtual slice reconstruction
- 2D damaged tissue repair
- Continuous spatial atlas generation
- Single-cell and spot-level spatial transcriptomics reconstruction
In our large-scale MERFISH mouse brain experiment, STITCH expands 54 observed slices into a continuous atlas of 571 slices and more than 11 million cells within approximately 5 hours on a single commodity GPU.
The datasets used in our paper are available from the following public resources:
- Stereo-seq Drosophila: Spateo Repository and direct data link
- MERFISH Mouse Brain: Allen Brain Cell Atlas
- Visium BRCA and DLPFC: So3D Database
- Xenium 2D Mouse Brain: 10x Genomics Portal
The full source code, examples, and reproducibility instructions will be released upon publication.
If you find STITCH useful for your research, please cite:
@article{Wang2026STITCH,
title = {STITCH: Spatial Transcriptomics Imputation via Flow Matching with Internal Learning},
author = {Wang, Sui and Wang, Xinyu and Peng, Qiangwei and Li, Tiejun},
journal = {bioRxiv},
year = {2026},
doi = {10.64898/2026.06.03.729557},
url = {https://www.biorxiv.org/content/early/2026/06/06/2026.06.03.729557}
}