GR2ST: spatial transcriptomics prediction based on graph-enhanced multimodal contrastive learning.

Area of Science:

• Computational biology
• Bioinformatics
• Genomics

Background:

• Spatial transcriptomics provides valuable gene expression and spatial coordinate data, crucial for disease mechanism research and cancer prognosis.
• High cost and time requirements of spatial transcriptomics limit its widespread application.
• Existing deep learning methods struggle to effectively integrate histology images with spatial transcriptomics data.

Purpose of the Study:

• To develop a deep learning model, GR2ST, for predicting spatial transcriptomics from histology images.
• To effectively integrate histological image features with spatial transcriptomic data for enhanced analysis.
• To advance the application of spatial transcriptomics in understanding disease mechanisms.

Main Methods:

• GR2ST utilizes a pre-trained pathology model for histological feature extraction.
• A dual-branch graph architecture with dynamic functional and spatial graphs captures complex tissue spot interactions.
• Multimodal contrastive learning aligns image and gene expression data, while a Cell-Type Guided Multi-Branch Regression Head adaptively generates gene expression.

Main Results:

• GR2ST demonstrates effectiveness and robustness on three cancer-related spatial transcriptomics datasets, including cutaneous squamous cell carcinoma and human breast cancer cohorts.
• The model successfully integrates histological features with gene expression data.
• The dual-branch graph architecture effectively models spot interactions in heterogeneous tissues.

Conclusions:

• GR2ST offers a powerful new method for predicting spatial transcriptomics from histology images.
• The model's ability to integrate multimodal data enhances the utility of spatial transcriptomics in biomedical research.
• GR2ST has the potential to reduce the cost and time associated with spatial transcriptomics sequencing.