Accurately Predicting Cell Type Abundance from Spatial Histology Image Through HPCell.

Area of Science:

• Computational Biology
• Genomics
• Pathology

Background:

• Spatial transcriptomics (ST) offers unparalleled insights into tissue architecture and cellular interactions.
• High sequencing costs limit the widespread adoption of ST.
• Existing methods often predict transcriptomic profiles but rarely estimate cell-type abundance directly from histology images.

Purpose of the Study:

• To develop a cost-effective deep learning framework, HPCell, for inferring cell-type abundance directly from H&E-stained histology images.
• To address the limitations of high sequencing costs in spatial transcriptomics.

Main Methods:

• HPCell utilizes a deep learning framework with pathology foundation, hypergraph, and Transformer modules.
• Whole-slide images (WSIs) are patched and processed for morphological features.
• A hypergraph models spatial relationships, and a Transformer captures long-range dependencies.

Main Results:

• HPCell accurately estimates cell-type abundance from histology images.
• The framework demonstrates superior performance compared to state-of-the-art methods across multiple ST datasets.
• HPCell provides a scalable and cost-effective solution for tissue analysis.

Conclusions:

• HPCell offers a viable alternative to expensive sequencing for cell-type abundance estimation in ST.
• This approach facilitates deeper understanding of tissue structure and cellular interactions.
• HPCell has the potential to significantly broaden the accessibility and application of spatial transcriptomics.