Bering: joint cell segmentation and annotation for spatial transcriptomics with transferred graph embeddings.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Single-cell spatial transcriptomics offers subcellular resolution for molecular mechanism insights.
• Accurate cell segmentation and annotation are critical but challenging, limiting downstream analysis.
• Existing methods often rely on cell staining, losing transcriptome depth and spatial patterns.

Purpose of the Study:

• To introduce Bering, a novel graph deep learning model for spatial transcriptomics.
• To leverage transcript colocalization for noise-aware cell segmentation and molecular annotation.
• To improve the accuracy and efficiency of spatial omics data analysis.

Main Methods:

• Developed Bering, a graph deep learning model utilizing transcript colocalization.
• Applied Bering to 2D and 3D spatial transcriptomics data.
• Benchmarked Bering against state-of-the-art methods.
• Constructed pre-trained models for transfer learning and self-distillation.

Main Results:

• Bering demonstrated superior cell segmentation accuracy compared to existing methods.
• The model successfully detected more transcripts across various technologies and tissues.
• Pre-trained Bering models achieved high segmentation accuracy on new datasets via transfer learning.

Conclusions:

• Bering effectively addresses challenges in spatial transcriptomics segmentation and annotation.
• The model enhances transcriptome depth and spatial colocalization pattern learning.
• Bering's capabilities advance the analysis of spatial omics data.