CosTaL

CosTaL transforms high-dimensional single-cell data into a weighted k-nearest-neighbor (kNN) graph and detects cellular communities using cosine similarity, Tanimoto coefficient refinement, and Leiden's algorithm.

Key Features:

• Exact kNN construction: Builds an exact k-nearest-neighbor graph using cosine similarity to capture initial relationships between cells.
• Tanimoto refinement: Re-weights graph edges using the Tanimoto coefficient to refine pairwise relatedness.
• Community detection: Applies Leiden's algorithm to identify graph partitions and cellular communities.
• Weighted kNN representation: Represents each cell as a vertex and encodes relatedness as weighted edges in the graph.
• Data modalities: Targets multidimensional single-cell datasets, including cytometry and single-cell RNA-sequencing (scRNA-seq) data.
• Benchmarking: Evaluated across seven benchmark cytometry datasets and six single-cell RNA-seq datasets using six evaluation metrics.
• Comparative performance: Demonstrates equivalent or higher clustering effectiveness compared with PhenoGraph, Scanpy, and PARC.
• Performance characteristics: Exhibits high efficiency on small datasets and acceptable scalability for larger datasets.

Scientific Applications:

• Single-cell clustering: Identification of cellular subpopulations in cytometry datasets.
• scRNA-seq analysis: Detection of transcriptionally defined cell communities in single-cell RNA-sequencing data.
• Community detection in high-dimensional data: Graph-based discovery of structure in multidimensional single-cell feature spaces.
• Method benchmarking: Comparative evaluation of clustering methods using standardized benchmark datasets and metrics.

Methodology:

Constructs an exact kNN graph using cosine similarity, re-weights edges with the Tanimoto coefficient to produce a weighted kNN graph, and applies Leiden's algorithm for community detection; performance evaluated on seven cytometry and six scRNA-seq datasets using six evaluation metrics.