A deep learning-based multiscale integration of spatial omics with tumor morphology.

Area of Science:

• Oncology
• Computational Biology
• Genomics

Background:

• Spatial transcriptomics (spTx) provides crucial insights into tumor microenvironments and therapeutic targets.
• Current spTx technologies are not yet widely accessible for routine clinical use.
• Hematoxylin and eosin (H&E) slides are standard in cancer diagnostics.

Purpose of the Study:

• To develop a deep learning method for integrating spTx data with tumor morphology from H&E slides.
• To enable prediction of spTx from routinely available H&E images.
• To bridge the gap between advanced spTx research and clinical practice.

Main Methods:

• A deep learning-based approach named MISO (multiscale integration of spTx with tumor morphology) was developed.
• MISO was trained to predict spTx data using H&E stained histological images.
• The model was validated on 72 10X Genomics Visium samples and 348 additional samples from the MOSAIC consortium.

Main Results:

• MISO successfully predicted spTx from H&E images across multiple cancer types.
• The method demonstrated superior performance compared to existing approaches in extensive benchmarks.
• MISO achieved near single-cell resolution for spatially-resolved gene expression prediction.

Conclusions:

• MISO offers a scalable solution for predicting spatial gene expression from H&E slides.
• This approach can democratize the use of spTx data in cancer research and diagnostics.
• MISO facilitates deeper understanding of tumor biology and identification of therapeutic targets.