BIDCell

BIDCell applies self-supervised deep learning to segment cells and assign transcripts in subcellular imaging transcriptomics by leveraging biologically-informed loss functions that relate spatially resolved gene expression to cell morphology.

Key Features:

• Biologically-Informed Learning: Integrates cell-type data, including single-cell transcriptomics from public repositories, with detailed cell morphology to inform loss functions.
• Self-Supervised Framework: Learns from spatial transcriptomics and imaging data without requiring extensive labeled datasets.
• Comprehensive Evaluation Metrics: Assesses performance using metrics across five complementary categories tailored for cell segmentation tasks.

Scientific Applications:

• Cross-Platform and Tissue Evaluation: Outperforms existing state-of-the-art methods across various tissue types and technology platforms according to multiple evaluation metrics.
• Single-Cell Spatial Expression Analysis: Improves cell segmentation and transcript assignment for single-cell spatial expression studies.
• Reduced Assignment Contamination: Mitigates fragmented or oversized cell segmentation errors that can lead to inaccurate expression assignments due to contamination.

Methodology:

Training is self-supervised: a deep learning model is trained using biologically-informed loss functions that relate spatially resolved gene expression data to cell morphology, mitigating fragmented or oversized cells and reducing assignment contamination.