PRTS: predicting single-cell spatial transcriptomics from histological images

Background

Spatial transcriptomics (ST) technologies reveal the spatial organization of gene expression in tissues, providing critical insights into development, neurobiology, and cancer. However, the high cost and technical complexity of ST limit its broad application, especially at single-cell resolution.

What is PRTS?

A team led by Prof. Fei Ling from South China University of Technology developed PRTS (Pathology-driven Reconstruction of Transcriptomic States), a deep learning framework that predicts single-cell-resolution spatial transcriptomics directly from H&E-stained histology images.

Input: Standard H&E images

Output: Single-cell × gene expression matrix (1,820 highly variable genes)

27x higher resolution compared to conventional ST spots

Trained and validated on mouse brain, human lung cancer, and breast cancer Visium HD datasets

Key Findings

Accurate Spatial Gene Prediction

PRTS reliably predicts the spatial expression patterns of key genes (e.g., Kcnma1, Plp1, Apoe) in mouse brain, matching ground truth ST data.

Single-Cell Annotation

The model identifies 21 cell subtypes (neurons, astrocytes, oligodendrocytes, etc.) with spatial distributions consistent with experimental data.

Robust Performance in Cancer Tissues

PRTS maintains prediction accuracy in human breast and lung cancer tissues, demonstrating generalizability to complex pathological environments.

Future Directions

Clinical Diagnostics: Convert routine H&E slides into transcriptomic maps for cancer subtyping and prognosis.

Drug Discovery: Identify spatially expressed genes in tumor microenvironments for target discovery.

Large-Scale Studies: Enable low-cost spatial transcriptomics for population-level research.

Cross-Platform Integration: Incorporate data from Xenium, Stereo-seq, etc., to build a universal model.

Sources: https://spj.science.org/doi/10.34133/research.0961