SwarmMAP: swarm learning for decentralized cell type annotation in single cell sequencing data.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell transcriptomic profiling generates large datasets, crucial for understanding tissue heterogeneity.
• Manual cell-type annotation is a bottleneck, suffering from poor reproducibility and scalability.
• Data privacy concerns hinder collaborative analysis of human single-cell datasets.

Purpose of the Study:

• To develop a standardized, automated, and privacy-preserving method for cell-type annotation.
• To leverage Swarm Learning for decentralized training of cell-type classification models.
• To evaluate SwarmMAP's performance and scalability across multiple tissue types.

Main Methods:

• Developed SwarmMAP, a framework applying Swarm Learning for decentralized machine learning.
• Trained cell-type classification models without raw data exchange between participating centers.
• Validated SwarmMAP on heart, lung, and breast single-cell RNA sequencing datasets.

Main Results:

• SwarmMAP achieved high F1-scores: 0.93 (heart), 0.98 (lung), and 0.88 (breast).
• Swarm Learning models demonstrated average performance of 0.907, comparable to centralized training.
• Increased dataset numbers enhanced prediction accuracy and broadened cell-type classification capabilities.

Conclusions:

• Swarm Learning offers an effective, automated solution for cell-type annotation in single-cell genomics.
• SwarmMAP addresses scalability and privacy challenges, facilitating broader data integration.
• The SwarmMAP framework promotes reproducible and scalable cell-type classification across research institutions.