scKSFD: federated distillation model with knowledge sharing for cell type classification of clinical transcriptome...

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) offers high-resolution cellular heterogeneity insights but faces collaboration hurdles due to privacy and data distribution issues.
• Existing federated learning methods often share model parameters, increasing privacy risks and not fully addressing scRNA-seq data's unique characteristics.

Purpose of the Study:

• To develop a privacy-preserving federated learning framework for cell type classification using scRNA-seq data.
• To address challenges in cross-institutional collaboration, including data privacy and distributional heterogeneity.

Main Methods:

• Proposed Federated Distillation framework with Knowledge Sharing (scKSFD) for privacy-preserving cell type classification.
• scKSFD aggregates knowledge in prediction space via probability-level soft label sharing on a reference dataset, reducing privacy risks.
• Integrated stratified proxy sampling and probability-level aggregation to handle rare cell populations and batch effects in scRNA-seq data.

Main Results:

• scKSFD achieved higher or comparable F1 scores against centralized and federated baselines across 42 clinical datasets under heterogeneous settings.
• Demonstrated statistically significant improvements in paired comparisons, highlighting robust performance.
• In a COVID-19 study, scKSFD improved classification over local training without sharing patient expression data.

Conclusions:

• scKSFD provides an effective federated distillation framework for multi-institutional single-cell transcriptomic analysis.
• The framework balances predictive performance, robustness, and data-sharing constraints.
• Enables collaborative analysis of sensitive scRNA-seq data while preserving patient privacy.