PSSD: Progressive Spatial-Semantic Decoupling for flow-based gene expression prediction from histology images

Predicting spatial gene expression from histology images offers a cost-effective complement to spatial transcriptomics. However, existing methods struggle to balance spatial continuity with functional heterogeneity, often producing over-smoothed predictions or neglecting spatial context.

We present PSSD, a conditional flow matching framework with progressive spatial-semantic decoupling. PSSD models spatial and semantic information through separate but interacting pathways and employs a three-stage architecture with decoupled flows, adaptive fusion, and cross-stream coupling to generate biologically coherent, high-fidelity gene expression profiles. Across seven spatial transcriptomics datasets spanning multiple tissues and resolutions, PSSD consistently achieved the highest Pearson correlation coefficients among the compared methods while better preserving biological boundaries and spatial autocorrelation. Under the same sampling protocol, PSSD reduced inference time from 32.04 to 3.98 min per sample on DLPFC compared with the diffusion-based Stem model and achieved approximately sevenfold acceleration on the BC and cSCC datasets without compromising predictive quality. These results demonstrate that flow-based spatial-semantic decoupling provides an effective and computationally efficient bridge between histology and transcriptomics.

The source code and data are available at https://github.com/ChyaZhang/PSSD.