Related Experiment Video
Updated: Jan 9, 2026
10:16
Mining Spatial Transcriptomics Datasets using DeepSpaceDB
Published on: September 5, 2025
624
Spatial transcriptomics expression prediction from histopathology based on cross-modal mask reconstruction and
Junzhuo Liu1, Markus Eckstein2, Zhixiang Wang3
1Faculty of Informatics and Data Science, University of Regensburg, Regensburg, Germany.
Medical Image Analysis
|December 2, 2025
Summary
This study introduces a deep learning method using contrastive learning to predict spatial gene expression from whole-slide images, overcoming data limitations in spatial transcriptomics for cancer research.
Area of Science:
- Computational biology
- Genomics
- Biomedical imaging
Background:
- Spatial transcriptomics provides crucial gene expression data for tumor microenvironment analysis and cancer diagnosis.
- Acquiring large-scale spatial transcriptomics data is challenging due to high costs.
- Existing methods struggle with limited spatial transcriptomics datasets.
Purpose of the Study:
- To develop a novel deep learning method for predicting spatially resolved gene expression from whole-slide images (WSIs).
- To establish a correspondence between histopathological morphology and spatial gene expression using multi-modal contrastive learning.
- To enhance feature-level fusion between modalities via cross-modal masked reconstruction.
Main Methods:
- A contrastive learning-based deep learning framework is proposed.
- Multi-modal contrastive learning aligns histopathological features with gene expression.
- Cross-modal masked reconstruction is employed as a pretext task for feature fusion.
- The method does not require large pretraining datasets or abstract semantic representations.
Main Results:
- The method accurately predicts spatially resolved gene expression from WSIs.
- Significant improvements in Pearson Correlation Coefficient (PCC) were observed for predicting highly expressed, highly variable, and marker genes (6.27%, 6.11%, and 11.26% increase, respectively).
- The approach preserves gene-gene correlations and is effective with limited sample datasets.
- Potential for cancer tissue localization based on biomarker expression was demonstrated.
Conclusions:
- The developed method offers an effective solution for predicting spatial gene expression from WSIs, particularly in data-limited scenarios.
- This approach enhances the utility of spatial transcriptomics in cancer research and clinical diagnosis.
- The method shows promise for advancing computational pathology and biomarker discovery.
