SegJointGene: joint cell segmentation and spatial gene prioritization by information entropy guided convolutional...

Area of Science:

• Computational biology
• Bioinformatics
• Genomics

Background:

• Accurate cell segmentation is crucial for spatial transcriptomics and proteomics.
• Nuclear staining alone is insufficient for precise cell boundary detection in dense tissues.
• Integrating molecular data for segmentation is computationally challenging.

Purpose of the Study:

• To develop a deep learning framework for joint cell segmentation and spatial gene prioritization.
• To improve the accuracy of assigning molecular signals to cell boundaries in complex tissues.
• To identify genes critical for cell-type-specific segmentation.

Main Methods:

• Developed SegJointGene, a deep learning framework integrating nuclei images with spatial gene/protein expression data.
• Utilized an information-entropy-guided convolutional neural network and an information discarding score.
• Iteratively refined gene prioritization and cell boundaries for convergent results.

Main Results:

• SegJointGene achieved 5-20% higher accuracy in assigning molecular signals to cell boundaries compared to existing methods.
• Demonstrated robust performance across various gene numbers and imaging resolutions.
• Prioritized genes were enriched in structural, developmental, and synaptic signaling pathways.

Conclusions:

• SegJointGene effectively integrates multi-omics spatial data for enhanced cell segmentation.
• The framework accurately identifies cell boundaries and prioritizes biologically relevant genes.
• This approach advances the study of molecular organization within complex tissue architectures.