GCNLA: Inferring Cell-Cell Interactions From Spatial Transcriptomics With Long Short-Term Memory and Graph...

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Cell-cell communication is crucial for physiological homeostasis and complex biological processes.
• Existing methods for identifying cell-cell interactions are limited by relying solely on neighboring cell gene expression and spatial data.
• Spatial transcriptomics offers advanced tools for investigating diverse biological tissues.

Purpose of the Study:

• To develop a novel network architecture, GCNLA (Graph Convolution Network and Long Short-Term Memory Attention module), for enhanced cell-cell interaction inference.
• To capture both spatial structures and interactions between distant cells, overcoming limitations of current methods.
• To reconstruct complete cell-cell interaction networks and enable downstream analyses like cell clustering.

Main Methods:

• Proposed GCNLA architecture integrating graph convolution, long short-term memory, and attention modules with residual connections.
• Utilized inner product decoding with cosine similarity for inferring cell-cell interactions.
• Applied GCNLA to seqFISH and MERFISH datasets for validation.

Main Results:

• GCNLA effectively learns spatial cell structures and captures interaction information between distal cells.
• The attention module enhances the extraction of cell-cell interaction features.
• Experimental results demonstrate GCNLA's superior robustness and noise immunity compared to existing methods.
• GCNLA successfully reconstructs complete cell-cell interaction networks.

Conclusions:

• GCNLA provides a robust and effective method for inferring cell-cell interactions from spatial transcriptomics data.
• The model's ability to capture long-range interactions and reconstruct networks advances the field.
• Learned features from GCNLA facilitate downstream analyses, including spatial cell clustering and resolving cellular heterogeneity.