BIDCell: Biologically-informed self-supervised learning for segmentation of subcellular spatial transcriptomics data.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Subcellular imaging transcriptomics offers high-resolution spatial gene expression mapping.
• Current analytical methods face challenges in accurate cell identification and transcript assignment, leading to segmentation errors.
• Fragmented or oversized cells in existing methods result in contaminated expression data.

Purpose of the Study:

• To introduce BIDCell, a novel self-supervised deep learning framework for enhanced cell segmentation in spatial transcriptomics.
• To address limitations in current cell segmentation methods by integrating gene expression and cell morphology.
• To improve the accuracy of transcript assignment and reduce data contamination in spatial expression analyses.

Main Methods:

• Developed BIDCell, a self-supervised deep learning framework utilizing biologically-informed loss functions.
• Integrated spatially resolved gene expression data with cell morphology information.
• Incorporated cell-type data, including single-cell transcriptomics from public repositories.

Main Results:

• BIDCell demonstrated superior performance in cell segmentation across various metrics.
• The framework outperformed state-of-the-art methods on diverse tissue types and technology platforms.
• BIDCell effectively learns relationships between gene expression and cell morphology for accurate segmentation.

Conclusions:

• BIDCell significantly enhances cell segmentation accuracy in spatial transcriptomics.
• The framework holds substantial potential for advancing biological discovery through improved spatial expression analyses.
• BIDCell offers a robust solution to persistent analytical challenges in the field.