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

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell spatial transcriptomics offers subcellular resolution for cell identification and mechanism understanding.
• Accurate cell segmentation and annotation are critical but challenging, limiting current insights.
• Existing methods relying on nuclei/cell body staining reduce transcriptome depth and spatial relationship learning.

Purpose of the Study:

• To introduce Bering, a graph deep learning model for joint, noise-aware cell segmentation and molecular annotation in spatial transcriptomics.
• To leverage transcript colocalization relationships for improved accuracy in 2D and 3D spatial data.
• To develop generalizable pre-trained models for streamlined segmentation via transfer learning.

Main Methods:

• Developed Bering, a graph deep learning model utilizing transcript colocalization for segmentation and annotation.
• Employed graph embeddings for cell annotation as multi-modal input to enhance cell segmentation.
• Benchmarked Bering against state-of-the-art methods on diverse spatial transcriptomics datasets.

Main Results:

• Bering demonstrated significant improvements in cell segmentation accuracy across various spatial technologies and tissues.
• The model increased the number of detected transcripts compared to existing methods.
• Pre-trained Bering models achieved high segmentation accuracy on new data through transfer learning and self-distillation.

Conclusions:

• Bering effectively addresses the challenge of cell segmentation and annotation in spatial transcriptomics by using transcript colocalization.
• The model's generalizability, shown through pre-trained models, facilitates broader application in spatial biology.
• Bering enhances transcriptome depth and spatial relationship analysis, advancing understanding of cellular mechanisms.