scBOL: a universal cell type identification framework for single-cell and spatial transcriptomics data.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell transcriptomic technologies enable high-throughput gene expression profiling.
• Accurate cell type identification is crucial for understanding tissue heterogeneity and cell states.
• Existing methods struggle with novel cell types and spatial transcriptomic data, often neglecting spatial organization and batch effects.

Purpose of the Study:

• To address limitations in current cell type identification methods for single-cell and spatial transcriptomics.
• To propose a universal cell type identification task that labels known cell types and clusters novel ones.
• To develop a versatile, end-to-end algorithm for accurate cell type annotation across diverse transcriptomic data.

Main Methods:

• Proposed a novel task: universal cell type identification for single-cell and spatial transcriptomics.
• Developed scBOL, an end-to-end algorithm based on Bipartite prototype alignment.
• Employed mutual nearest cluster identification, cycle-consistent semantic anchor mining, and neighbor-aware prototypical learning.

Main Results:

• scBOL effectively identifies mutual nearest clusters as common cell types and mines semantic anchors for structural association.
• The neighbor-aware prototypical learning paradigm enhances inter-cluster separability and intra-cluster compactness.
• Extensive benchmarks demonstrate scBOL's superiority over state-of-the-art methods in both non-spatial and spatial transcriptomics.

Conclusions:

• scBOL provides a comprehensive algorithmic framework for universal cell type identification across varied single-cell data types.
• The method successfully aligns known cell types and separates novel cell types, outperforming existing approaches.
• scBOL is implemented in Python (Pytorch) and publicly available, facilitating its application in transcriptomic data analysis.