Deep Learning-Enabled Integration of Histology and Transcriptomics for Tissue Spatial Profile Analysis.

Area of Science:

• Computational Biology
• Genomics
• Pathology

Background:

• Spatially resolved transcriptomics (SCST) offers subcellular gene expression data within tissue context.
• Current deep learning methods for SCST primarily focus on sequence and spatial data, often overlooking histopathology.
• Integrating histological images with transcriptomic data can provide richer biological insights.

Purpose of the Study:

• To introduce GIST, a novel deep learning framework for integrating gene expression and histology data for spatial cellular profiling.
• To leverage histopathology foundation models and a hybrid graph transformer for enhanced feature extraction and integration.
• To improve the accuracy of spatial domain identification and microenvironment segmentation in SCST analysis.

Main Methods:

• Utilized pre-trained histopathology foundation models for feature extraction from histology images.
• Developed a hybrid graph transformer model to integrate transcriptome features with histological features.
• Applied GIST to human lung, breast, and colorectal cancer datasets.

Main Results:

• GIST effectively revealed distinct spatial domains within cancer tissues.
• Significantly improved the accuracy of microenvironment segmentation by denoising transcriptomics data.
• Outperformed existing deep learning methods, achieving up to 49.72% improvement in analysis accuracy.
• Enabled more precise gene expression analysis and identification of prognostic marker genes.

Conclusions:

• GIST provides a generalizable framework for integrating histology and spatial transcriptome data.
• The integration enhances the understanding of spatial organization and functional dynamics in the tumor microenvironment.
• This approach offers novel insights for cancer research and biomarker discovery.