FineST: contrastive learning integrates histology and spatial transcriptomics for nuclei-resolved ligand-receptor...

Area of Science:

• Computational Biology
• Genomics
• Cancer Research

Background:

• Spatial transcriptomics (ST) is crucial for understanding cell-cell communication (CCC) in development and disease.
• Current ST methods face limitations in resolution and data sparsity, hindering detailed CCC pattern analysis in tissues.

Purpose of the Study:

• To introduce FineST, a deep contrastive learning model for fine-grained spatial transcriptomics analysis.
• To fuse ST data with histology images for enhanced resolution and CCC pattern discovery.

Main Methods:

• Developed FineST, a deep contrastive learning model integrating a histology foundation model with ST data.
• Applied FineST to colorectal cancer VisiumHD and breast cancer Xenium datasets.
• Evaluated FineST for nuclei segmentation, RNA expression imputation, cell type prediction, and CCC pattern identification.

Main Results:

• FineST significantly improves high-resolution RNA imputation, cell type prediction, and CCC pattern discovery compared to existing methods.
• The model enables precise nuclei segmentation and identification of intricate ligand-receptor interactions.
• Analysis of cancer datasets revealed novel insights into tumor-immune interactions, including invasive fronts, tertiary lymphoid structures, and therapy resistance barriers.

Conclusions:

• FineST offers a new paradigm for ST analysis by integrating histology images, overcoming resolution and sparsity limitations.
• This approach enhances the understanding of complex biological processes and disease mechanisms, particularly in cancer.
• The integration of histology images with ST data provides a powerful tool for biological discovery.