PSSD: Progressive Spatial-Semantic Decoupling for Flow-Based Gene Expression Prediction from Histology Images

Area of Science:

• Computational biology
• Bioinformatics
• Genomics

Background:

• Predicting spatial gene expression from histology images is a cost-effective alternative to spatial transcriptomics.
• Existing methods face challenges in balancing spatial continuity with functional heterogeneity, leading to over-smoothed predictions or ignored spatial context.

Purpose of the Study:

• To develop a novel framework for accurate and efficient prediction of spatial gene expression from histology images.
• To address limitations of existing methods in preserving spatial continuity and functional heterogeneity.

Main Methods:

• Introduced PSSD (Progressive Spatial-Semantic Decoupling), a conditional flow matching framework.
• Employed a three-stage architecture with decoupled flows, adaptive fusion, and cross-stream coupling.
• Modeled spatial and semantic information through separate, interacting pathways.

Main Results:

• PSSD achieved the highest Pearson correlation coefficients across seven spatial transcriptomics datasets.
• Demonstrated improved preservation of biological boundaries and spatial autocorrelation compared to existing methods.
• Significantly reduced inference time (e.g., from 32.04 to 3.98 min/sample on DLPFC) with a sevenfold acceleration on other datasets without compromising predictive quality.

Conclusions:

• Flow-based spatial-semantic decoupling provides an effective and computationally efficient bridge between histology and transcriptomics.
• PSSD generates biologically coherent, high-fidelity gene expression profiles.
• The framework offers a promising solution for cost-effective spatial gene expression prediction.