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

Axel Faes, Stéphanie M. van den Berg, Maryam Amir Haeri · 2026

Read Privacy-preserving federated tensor decomposition of single-cell immune data: recovering multicellular programs across institutions by Axel Faes, Stéphanie M. van den Berg, Maryam Amir Haeri (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-24938.txt.

Introduction

Axel Faes, Stéphanie van den Berg, and Maryam Amir Haeri find shared gene programmes in immune data without pooling patient cells. A tensor decomposition of donor by cell type by gene data can reveal multicellular programmes: patterns of gene activity that span cell types and separate sick donors from healthy ones. Privacy rules increasingly block that pooling. Their federated estimator instead leaves the cells at each hospital.

Each site computes a small local subspace that captures its share of the programme. A central merge then stacks these subspaces and combines them, using a global mean centering done across sites. The authors prove the merged result matches the pooled decomposition, up to truncation. On a 261 donor lupus atlas, the method recovers the known interferon programme. Its scores for gene enrichment and for telling cases from controls nearly match the pooled analysis, and no donor level data ever leaves a site.

Catalog record

CREATORAxel Faes, Stéphanie M. van den Berg, Maryam Amir Haeri
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE43 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.24938

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Privacy-preserving federated tensor decomposition of single-cell immune data: recovering multicellular programs across institutions Axel Faes, Stéphanie M. van den Berg, Maryam Amir Haeri Abstract Tensor decomposition of donor × \times cell-type × \times gene single-cell data recovers multicellular programs : coordinated axes of inter-individual transcriptional variation that span cell types and stratify disease. Such analyses assume all cells are centralised, yet immune single-cell atlases are increasingly multi-institution, multi-ancestry, and governed: patient cells often cannot be pooled. We present a federated estimator for multicellular tensor decomposition. Each site computes a low-dimensional local program subspace; a coordinator merges these by stacked singular value decomposition under federated global-mean centering, an operation provably equivalent (up to truncation) to the centralised decomposition. This centering also makes the merge robust to site-label confounding (program AUC held flat at 0.957 0.957 across case-fraction skew, whereas naive per-site centering degrades to 0.861 0.861 ). Only program subspaces (never cells or donor-level scores) leave a site, and aggregation is a summation compatible with secure aggregation. On a 261-donor systemic lupus erythematosus (SLE) atlas the federated estimator recovers the canonical interferon multicellular program (gene-mode ISG enrichment AUC 0.998 0.998 ; case–control separation 0.958 0.958 ; bootstrap Δ ​ AUC = − 0.000 \Delta\text{AUC}=-0.000 , 95% CI [ − 0.004 , + 0.012 ] [-0.004,+0.012] vs. centralised; label-free cross-validated AUC 0.960 0.960 ), faithfully across institution-scale and multi-ancestry partitions, and across the three real collection sites of a COVID-19 atlas (subspace correlation 0.989 0.989 ). We further show the estimator recovers the full multicellular program when no site observes all cell types (subspace correlation 1.000 1.000 , exact by construction), a setting fixed-feature federated PCA cannot address; it is private at the complete-panel level when the donor coupling is computed under secure computation and only loadings are released. On an interstitial-lung-disease atlas, where the fibrotic response is intrinsically multicellular, the recovered program predicts disease better than the best single cell type (AUC 0.96 0.96 vs. 0.91 0.91 ; gap 95 % 95\% CI excludes zero, robust across seeds) and the advantage survives federation, the regime the method targets; a smaller fibrosing-cholangiopathy liver cohort is consistent (label-specific by permutation, p = 0.005 p=0.005 ); on compartment-dominated blood, gut, kidney and heart cohorts it matches but does not exceed the best single compartment. A membership-inference evaluation quantifies the privacy gain of secure aggregation (attack AUC reduced from 0.91 0.91 to 0.61 0.61 for the complete-panel scheme), directly addressing the donor re-identification risk recently demonstrated for single-cell count matrices. The method enables cross-institution and cross-ancestry recovery and comparison of multicellular immune programs without sharing cells. I Introduction Population-scale single-cell RNA sequencing has made it possible to study how the coordinated state of the immune system varies between individuals. An established framing represents a patient cohort as a third-order tensor of donor \times cell-type \times gene pseudobulk expression and decomposes it to recover multicellular programs : factors whose loadings span several cell types and capture inter-individual axes of variation that no single cell type expresses in isolation. Interpretable tensor decomposition (scITD) introduced this view and used it to stratify lupus and COVID-19 patients [ 29 ] ; related tensor methods model cell–cell communication across contexts [ 4 ] and condition-resolved variation [ 10 ] . These methods are centralised : they assume every cell is available in one place. Such programs are of direct biomedical interest: they place each patient on continuous axes of immune state, stratify cohorts beyond case/control labels, and link coordinated transcriptional shifts to disease severity, treatment response, and genetics. Their value grows with cohort size and diversity, so the most informative analyses are precisely those spanning many institutions and ancestries, the setting hardest to centralise. In biomedical practice that assumption rarely holds. Immune single-cell atlases are assembled across hospitals, biobanks, and national cohorts [ 35 ] , and the underlying data are governed: patient-derived single-cell profiles are subject to consent and data-protection constraints that frequently preclude pooling raw cells [ 36 , 38 ] . Multi-ancestry comparisons compound the problem, since the most informative contrasts often span populations held by different consortia. Federated analysis (computing over data that remain at their source) is the natural response [ 28 , 25 ] , and federated methods now exist for single-cell batch correction [ 5 ] , cell-type classification under homomorphic encryption [ 37 ] , and quantitative-trait-locus mapping [ 32 ] . None recovers multicellular programs : the tensor-decomposition task that scITD addresses centrally has no federated counterpart. We close that gap. Our contributions are deliberately scoped and reported with their limits: 1. Vertical-FL recovery of cross-panel multicellular programs : recovery of the full multicellular program when no site observes all cell types , by coupling through the shared donor mode. This is a vertical federated setting (shared donors, partitioned cell-type panels); it recovers programs spanning cell types no single site sees, which fixed-feature federated PCA cannot handle at all. 2. A confounding-robust federated estimator : sites share only a low-dimensional program subspace; the coordinator merges local subspaces by stacked SVD under federated global-mean centering, reproducing the centralised decomposition up to truncation, and the aggregation is a summation hence compatible with secure aggregation [ 6 ] . This complete-panel merge instantiates distributed PCA [ 18 , 20 ] ; the methodological additions are that the global-mean centering makes the merge robust to site-label confounding (Table I ) and, with contribution 1, the cross-panel donor coupling. 3. A membership-inference privacy evaluation [ 39 ] quantifying the leakage of each release granularity and the protection conferred by secure aggregation, the practical guarantee on which our privacy claims rest (formal differential privacy [ 15 ] is out of reach at present single-cell cohort sizes; Section V ). 4. A cross-tissue map of when multicellular recovery matters (§ IV-E ): across seven cohorts the multicellular program out-predicts the best single cell type only where cross-cell-type coordination is strong, with interstitial lung disease the stated result (cross-validated AUC 0.959 vs. 0.910 , gap 95\% CI excluding zero); a smaller fibrosing-cholangiopathy liver cohort corroborates (separation label-specific by permutation, p=0.005 ), and the program is at parity on compartment-dominated blood, gut, kidney and heart, establishing when the method earns its cost. We are explicit about scope (Section V ): multicellular decomposition out-predicts the best single cell type only where cross-cell-type coordination is strong (interstitial lung disease and fibrosing-cholangiopathy liver) and is at parity on compartment-dominated blood, gut, kidney and heart; the central contribution is faithful private federated recovery of established coordinated programs, and formal differential privacy is left to future work. II Related Work Centralised multicellular decomposition. scITD [ 29 ] factorises a donor \times cell-type \times gene tensor by Tucker decomposition to recover multicellular programs; Tensor-cell2cell [ 4 ] models cell–cell communication across contexts and PARAFAC2-RISE [ 10 ] condition-resolved variation, while MOFA [ 3 ] integrates modalities by matrix factorisation. DIALOGUE [ 24 ] recovers multicellular programs by cross-cell-type association rather than tensor decomposition. All operate on pooled data. Federated single-cell analysis. Existing federated single-cell methods target different tasks: FedscGen [ 5 ] performs batch-effect correction, PriCell [ 37 ] cell-type classification under homomorphic encryption, and foundation-model pretraining has been federated as well [ 44 ] .

Read the complete text