Privacy-preserving federated tensor decomposition of single-cell immune data: recovering multicellular programs across institutions

Researchers developed a federated tensor decomposition method that enables privacy-preserving analysis of single-cell immune data across multiple institutions without sharing raw patient data. The approach recovers multicellular immune programs—coordinated patterns of gene expression across cell types—while protecting patient privacy through secure aggregation, demonstrated on systemic lupus erythematosus and COVID-19 datasets.

This research addresses a critical bottleneck in collaborative biomedical research: the tension between advancing scientific understanding through multi-institutional data integration and protecting patient privacy under increasingly stringent governance requirements. Traditional approaches require centralizing sensitive single-cell data, creating regulatory and ethical barriers that prevent many institutions from participating in large-scale studies.

The federated approach fundamentally changes this dynamic by ensuring only aggregated program subspaces—not raw cellular data—leave each site. The method's mathematical soundness is particularly notable: the authors prove their stacked SVD approach with global-mean centering produces results equivalent to centralized analysis while remaining robust to site-specific biases. Testing on a 261-donor lupus atlas achieved performance metrics nearly identical to centralized methods (AUC 0.958 vs 0.957), while membership-inference attacks show secure aggregation substantially reduces privacy risks.

The practical implications extend beyond privacy compliance. The method handles incomplete cell-type coverage across sites—a common real-world constraint that defeats conventional federated PCA—enabling discovery of cross-institutional immune programs invisible to single-site analyses. Validation across three real COVID-19 sites and multiple disease cohorts demonstrates reproducibility and clinical relevance.

For the biotech and healthcare AI sectors, this work reduces institutional barriers to collaborative research, potentially accelerating discovery in immunology and disease stratification. The demonstrable privacy protections could influence how regulatory bodies approach data governance, encouraging larger consortiums. As multi-institutional atlases become standard research infrastructure, federation-friendly methods shift from optional to essential, creating demand for similar privacy-preserving approaches across genomics and beyond.

• →Federated tensor decomposition recovers multicellular immune programs without sharing raw patient data across institutions
• →Method achieves performance equivalent to centralized analysis while reducing privacy attack success from 91% to 61%
• →Approach handles incomplete cell-type coverage across sites, enabling discovery impossible with fixed-feature federated methods
• →Validation on lupus and COVID-19 cohorts demonstrates clinical utility and cross-ancestry reproducibility
• →Results suggest privacy-preserving methods could accelerate large-scale biomedical collaboration by removing institutional data-sharing barriers