Related Experiment Video
Updated: Jul 10, 2026
10:16
Mining Spatial Transcriptomics Datasets using DeepSpaceDB
Published on: September 5, 2025
DFCE-KanT: Predicting Spatial Gene Expression from Histology Images via Contrastive Learning
Fang Li1, Pengyu Wang1, Junjie Shen2,1
1Department of Health Statistics, College of Preventive Medicine, Army Medical University, Chongqing 400038, China.
Bioinformatics (Oxford, England)
|July 9, 2026
Summary
We developed DFCE-KanT, a cost-effective method to predict spatial gene expression from H&E images. This approach effectively utilizes spatial context, outperforming existing methods for enhanced biological insights.
Area of Science:
- Computational biology
- Bioinformatics
- Genomics
Background:
- Spatial transcriptomics (ST) provides detailed molecular profiles but is expensive for large-scale application.
- Hematoxylin and eosin (H&E) stained whole slide images (WSIs) offer a low-cost alternative for inferring spatial gene expression.
- Current deep learning methods often underutilize the spatial context present in H&E images.
Purpose of the Study:
- To develop a novel deep learning framework, DFCE-KanT, for accurate spatial gene expression prediction from H&E images.
- To effectively integrate tissue image features, gene expression data, and spatial location information.
- To address the limitations of existing methods in leveraging spatial context from H&E WSIs.
Main Methods:
- DFCE-KanT employs a contrastive learning framework.
- Feature extraction from H&E images is enhanced using DenseNet with a Feature Channel Enhancement (FCE) attention mechanism.
- A Kolmogorov-Arnold Network (KAN) combined with Transformer multi-head attention (KanT) automatically learns optimal fusion strategies for spatial information via positional encoding.
Main Results:
- DFCE-KanT demonstrates superior performance in spatial gene expression prediction compared to existing methods across five diverse datasets (HER2+, cSCC, Alex, HBC, and Liver).
- The KanT component enhances the model's ability to capture complex data patterns and nonlinear interactions.
- The framework effectively promotes compatibility between image and gene expression features in a shared embedding space.
Conclusions:
- DFCE-KanT offers a highly effective and computationally efficient solution for predicting spatial gene expression from routine H&E images.
- The method advances the potential of low-cost tissue imaging for large-scale spatial omics studies.
- The findings validate the efficacy of integrating enhanced image features with advanced fusion strategies for accurate spatial transcriptomics inference.
