CosTaL: an accurate and scalable graph-based clustering algorithm for high-dimensional single-cell data analysis.

Area of Science:

• Computational Biology
• Bioinformatics
• Data Science

Background:

• Analyzing large-scale, high-dimensional single-cell data presents significant computational challenges.
• Existing graph-based clustering methods may struggle with the complexity and size of modern single-cell datasets.

Purpose of the Study:

• To develop and evaluate a novel graph-based clustering method, Cosine-based Tanimoto similarity-refined graph for community detection (CosTaL).
• To improve the accuracy and scalability of community detection in large, multidimensional single-cell datasets.

Main Methods:

• CosTaL transforms single-cell data into a weighted k-nearest-neighbor (kNN) graph.
• It employs cosine similarity for exact kNN graph construction.
• The Tanimoto coefficient is utilized as a refining strategy to re-weight graph edges, enhancing clustering effectiveness.

Main Results:

• CosTaL demonstrated equivalent or superior effectiveness scores across seven cytometry and six single-cell RNA-sequencing benchmark datasets.
• Performance was evaluated using six distinct metrics, consistently outperforming state-of-the-art methods like PhenoGraph, Scanpy, and PARC.
• CosTaL exhibits high efficiency for small datasets and acceptable scalability for large datasets.

Conclusions:

• CosTaL offers a robust and effective solution for community detection in large-scale single-cell data analysis.
• The method's combination of cosine similarity and Tanimoto refinement provides improved clustering accuracy.
• Its efficiency and scalability make it a valuable tool for contemporary biological research.