RAG: a regularized adaptive graph-based method for rare-cell identification from single-cell expression data

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Rare-cell identification is crucial for understanding disease and development.
• Current methods struggle with varying cell densities, often misclassifying rare cells.

Purpose of the Study:

• To introduce the RAG method for improved rare-cell identification.
• To address limitations of existing neighborhood graph-based approaches.

Main Methods:

• Constructing a regularized adaptive graph using cell-specific radii and hybrid dissimilarity.
• Estimating locally scaled hybrid affinities for density normalization.

Main Results:

• RAG significantly outperformed six state-of-the-art methods on 10 single-cell RNA sequencing datasets.
• Achieved substantial improvements in precision, F1 score, and rare-type coverage.
• Demonstrated superior accuracy in a colorectal tumor dataset and identified known cell populations in other datasets.

Conclusions:

• RAG offers a robust and accurate method for rare-cell identification.
• The approach effectively handles varying cell densities and improves biological discovery.