STITCH: Spatial Transcriptomics Imputation via Flow Matching with Internal Learning

Abstract

AO_SCPLOWBSTRACTC_SCPLOWSpatial transcriptomics datasets frequently suffer from spatial gaps and missing regions due to sectioning artifacts, tissue damage, and the high cost of sequencing that limits tissue coverage. We present STITCH, a scalable and robust generative framework for multidimensional virtual spatial transcriptomics reconstruction. STITCH models intrinsic spatial-transcriptomic patterns directly from individual tissue samples, enabling reconstruction without requiring external reference atlases or matched histological image priors. The framework adopts a decoupled architecture that separates spatial morphology restoration from transcriptomic generation. STITCH first compresses high dimensional transcriptomic profiles into a low-dimensional latent representation through a spatial-aware graph autoencoder. For 3D cross-slice gaps, STITCH employs optimal transport-conditioned flow matching for spatial reconstruction, whereas 2D in-slice damage is repaired through an internal learning strategy. To generate the corresponding transcriptomic profiles, STITCH further establishes a point-wise conditional flow matching model in the latent space. This module achieves linear computational complexity, enabling continuous 3D atlas reconstruction of over 11 million cells within 5 hours on a single commodity GPU. Extensive evaluations across diverse spatial transcriptomics platforms, spanning both single-cell and spot-level technologies, demonstrate that STITCH consistently preserves transcriptomic identities, spatial topologies, and anatomical continuity. Overall, STITCH provides a scalable and platform-compatible computational framework for reconstructing high-resolution continuous spatial transcriptomic atlases.