STFormer: Learning to Explore Spot Relationships for Spatial Transcriptomics Prediction from Histology of Colorectal...

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomics (ST) provides crucial gene expression data within tumor microenvironments.
• High costs currently limit the widespread adoption of ST.
• Predicting ST from histology images offers a more accessible alternative.

Purpose of the Study:

• To develop a cost-effective and accurate method for predicting spatial transcriptomics from histology images.
• To overcome limitations of existing methods in handling stain variations and inter-spot correlations.

Main Methods:

• Introduction of STFormer, a deep learning model for ST prediction from Whole Slide Images (WSIs).
• Incorporation of a Style-Aug module for enhanced feature generalization via style transfer.
• Utilization of a Cross-WSI Transformer module to capture relationships between spots across different WSIs.

Main Results:

• STFormer demonstrated superior performance compared to existing methods (STNet, HistoGene, Hist2ST).
• Experiments on internal and external datasets confirmed the model's effectiveness.
• The model successfully addresses stain variation and improves inter-spot correlation analysis.

Conclusions:

• STFormer presents a significant advancement in predicting spatial transcriptomics from histology.
• The model offers a powerful and cost-effective solution for broader ST application in research.
• This approach has the potential to accelerate discoveries in tumor microenvironment research.