MCGAE: unraveling tumor invasion through integrated multimodal spatial transcriptomics.

Area of Science:

• Biomedical research
• Computational biology
• Genomics

Background:

• Spatially Resolved Transcriptomics (SRT) is crucial for understanding tissue microenvironments.
• Integrating multimodal data (gene expression, spatial, morphological) for spatial domain identification presents challenges.

Purpose of the Study:

• To present the Multi-view Contrastive Graph Autoencoder (MCGAE), a deep learning framework for spatial transcriptomics (ST) data analysis.
• To enhance spatial domain identification by integrating multimodal data.

Main Methods:

• Developed MCGAE, a framework using multi-view representations from gene expression and spatial adjacency matrices.
• Employed modular modeling, contrastive graph convolutional networks, and attention mechanisms.
• Integrated morphological image features for multimodal data processing.

Main Results:

• MCGAE demonstrated superior performance in spatial domain detection, data denoising, and trajectory inference on simulated and real SRT datasets.
• Outperformed existing methods in key ST analysis tasks.
• Successfully integrated histological and gene expression data in colorectal cancer liver metastases to identify tumor invasion regions.

Conclusions:

• MCGAE offers a powerful new tool for analyzing complex spatial transcriptomics data.
• The framework advances ST analysis, providing new capabilities for cancer and complex disease research.