Related Experiment Video
Updated: Jan 10, 2026
09:19
Spatial Profiling of Protein and RNA Expression in Tissue: An Approach to Fine-Tune Virtual Microdissection
Published on: July 6, 2022
5.3K
SciSt: single-cell reference-informed spatial gene expression prediction from pathological images
Yixin Li1, Fan Zhong2, Lei Liu2,3,4
1Institutes of Biomedical Sciences, Fudan University, 130 Dong'an Road, Xuhui District, Shanghai 200032, China.
Briefings in Bioinformatics
|November 20, 2025
Summary
This study introduces SciSt, a deep learning model that predicts spatial gene expression from H&E images. SciSt leverages pathological features and biological data for cost-effective, large-scale spatial analysis in disease research.
Area of Science:
- Computational Biology
- Bioinformatics
- Digital Pathology
Background:
- Spatial transcriptomics is crucial for understanding disease mechanisms but limited by sample scarcity and high costs.
- Clinical histopathology images (H&E-stained) offer a vast, cost-effective resource for spatial analysis.
- Existing methods struggle to accurately predict spatial gene expression from histopathology images due to limitations in capturing transcriptomic structures.
Purpose of the Study:
- To develop a novel deep learning framework, SciSt, for predicting spatial gene expression from histopathological images.
- To enhance biological interpretability and accuracy by integrating pathological features with biologically informed initial gene expressions.
- To enable efficient and cost-effective large-scale spatial analysis using readily available clinical image archives.
Main Methods:
- Developed SciSt, a deep learning framework integrating pathological features with biologically informed initial gene expressions.
- Generated initial gene expressions using a weighted strategy combining cell segmentation and single-cell reference data.
- Validated SciSt on three benchmark datasets and TCGA-BRCA and TCGA-LIHC cohorts for performance and generalization.
Main Results:
- SciSt achieved state-of-the-art performance in predicting spatial gene expression from H&E images.
- Outperformed existing models by 21.4% and 13.7% on benchmark datasets.
- Demonstrated robust generalization capabilities on clinical TCGA cohorts, highlighting its real-world applicability.
Conclusions:
- SciSt effectively predicts spatial gene expression from histopathological images, overcoming limitations of current methods.
- The framework enables cross-modal translation between morphology and gene expression, unlocking potential in clinical image archives.
- Integrating prior biological knowledge significantly improves the interpretability and scalability of biomedical AI models.
