Plant Spatio-Temporal Integration Network (PSTN): A framework for dynamic single-cell and spatial transcriptomics in fungal–plant interactions

In plain English

The framework learns cell-to-space mappings while encouraging neighboring time points to agree, aiming to preserve both stage-specific biology and temporal continuity during rice-blast infection.

What they found

• At 4,000 highly variable genes, reported mean squared error stayed below 0.09 at all stages.
• An independent mouse-cortex benchmark supported only the static reconstruction component across systems.

The catch

• The cross-system benchmark tests static reconstruction, not the full temporal claim.
• Results remain limited to the datasets described in the accepted manuscript.

Methods at a glance

Study type

Cross-modal integration method

Biological system

Rice infected with Magnaporthe oryzae

Organism

Rice and mouse cortex benchmark

Dataset

Matched rice scRNA-seq and spatial transcriptomics at 0, 12, and 24 hours

Sample size

Not reported

Model architecture

Expression reconstruction with bidirectional maximum-similarity temporal regularization

Training objective

Stage-specific cell-to-space mapping with temporal consistency

Code

PSTN code is available from the linked GitHub repository.

Data

The publisher record links a Zenodo archive for the repository.

Tasks and metrics

Tasks: Spatial expression reconstruction; Temporal-pattern recovery

Metrics: Mean squared error; Gene-wise correlation; Spot-wise correlation

Baselines: Tangram; SpaGE; cell2location; DestVI

Principal findings

• At 4,000 highly variable genes, mean squared error remained below 0.09 at every reported stage.

What remains unresolved

External validation: Mouse-cortex data supported the static reconstruction component.

• The temporal component was evaluated in one host-pathogen system.
• Cross-system validation covered static reconstruction rather than temporal integration.