REVIEW 3 major objections 5 minor 2 references
Batch effects can impair federated learning in multi-center omics studies
T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash
Pith's one-line read FedRBE, a privacy-preserving federated implementation of limma's removeBatchEffect(), matches centralized results to within 3.6E-13 on real and simulated omics data.
desk verdict A useful federated batch-correction tool with solid empirical equivalence checks, but the abstract overpromises and the missing-value equivalence is asserted rather than proven. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing machinery is a federated linear regression built on aggregating local cross-product matrices $X_i^T X_i$ and $X_i^T y_j$ instead of raw data, following the secure-regression approach of Karr et al. Because the existing federated QR decomposition lacks column pivoting, fedRBE adds a per-feature global mask that removes whole batch columns from the design matrix when a feature is absent from a batch, and redefines the reference batch when the original reference lacks the feature. This mask is what lets the method handle missing values and non-identical feature sets while mimicking limma's row-wise deletion and QR-based least-squares fit. Additive secret sharing (splitting each shared matrix into random pieces that sum to the original) is layered on top to hide individual contributions during aggregation.
What would settle it
Construct a distributed dataset with a feature missing from the reference batch and present in a non-contiguous subset of other batches, run fedRBE and centralized limma, and check whether any corrected value differs by more than about $10^{-10}$; a larger difference would refute the claimed mathematical equivalence.
Extended reading notes
Core claim
The central claim is that fedRBE is a drop-in federated substitute for limma::removeBatchEffect(): it receives results mathematically equivalent to the centralized result by computing, per feature, the same linear-regression adjustment from aggregated local matrices. Each client computes local $X^T X$ and $X^T Y$ for its samples; the coordinator sums them across clients, applies a per-feature mask that drops batches where the feature is absent (re-selecting the reference batch when needed), and solves for the regression coefficients that are then used to subtract the batch effect locally. Because only the aggregated cross-product matrices are exchanged, and because additive secret sharing masks those matrices in transit, patient-level data stay local. The paper's empirical comparison shows maximum absolute differences of $2.2\times10^{-13}$ on real datasets without SMPC and up to $3.6\times10^{-13}$ across simulations.
Load-bearing premise
The claim that the per-feature masking exactly reproduces limma's row-wise deletion and QR-based fit for every possible missingness pattern is asserted mathematically but only tested on three real datasets and simulations; unusual patterns could yield different corrected values.
Editorial extensions
If this is right
- Multi-center omics studies can correct batch effects without a central data repository, so downstream analyses like clustering or co-expression can run federated end-to-end.
- Proteomics and metabolomics data, which commonly contain missing values, become eligible for federated batch-effect correction without prior imputation.
- Because fedRBE is mathematically equivalent to limma, results from federated analyses can be compared directly with historical centralized limma-corrected results.
- The SMPC component protects intermediate computations from reconstruction attacks, addressing a known vulnerability of earlier distributed ComBat implementations.
Reading between the lines
- If the equivalence holds for arbitrary missingness patterns, a similar aggregation strategy could be applied to other linear-model-based harmonization methods, reusing the same cross-product aggregation.
- The per-feature mask's replacement of missing batch columns suggests that batches with very few observed features will have their effects estimated from a shrinking subset of samples; future work could quantify the minimal feature presence required for stable correction.
- Because the app is deployed on a federated platform, in practice the bottleneck may shift from the math to governance: deciding which features and covariates are hashed and shared, and whether the union-of-features policy leaks information about which centers measure which analytes.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript presents fedRBE, a federated implementation of limma's removeBatchEffect() built on the FeatureCloud platform, augmented with additive-secret-sharing SMPC and designed to accept missing values and non-identical feature sets across clients. The authors evaluate fedRBE on a set of simulations and on three real omics datasets (E. coli mass-spectrometry proteomics, ovarian-cancer microarray transcriptomics, and colorectal-cancer shotgun metagenomics), reporting maximum absolute differences on the order of 1e-13 relative to centralized limma. The abstract additionally claims a systematic assessment of how uncorrected batch effects impair federated k-means clustering and federated random forest classification on four datasets including metabolomics, but that assessment does not appear in the main text.
Significance. As a tool paper, fedRBE is a useful contribution: a drop-in federated replacement for a widely used batch-effect correction routine would enable privacy-preserving multi-center omics workflows, and the empirical validation against the external limma standard is a genuine strength. The numerical equivalence on the tested data is well supported, and the provision of code, app, and public datasets is commendable. The broader claim in the title and abstract about systematically quantifying the impact of batch effects on federated k-means and random forest is not supported by the presented experiments, so the significance of the paper as a study of batch effects in FL is currently much weaker than the framing suggests.
major comments (3)
- [Abstract and Title] The abstract and title promise a systematic assessment of uncorrected batch effects on federated k-means clustering and federated random forest classification using four multi-center omics datasets, including metabolomic data. The main text contains no such assessment: Table 2 lists three datasets, none of them metabolomic, and the Results section only evaluates the numerical equivalence of fedRBE to centralized limma. This mismatch is load-bearing for the paper's stated contribution; the authors must either include the promised FL experiments or revise the title, abstract, and framing to describe a tool paper on fedRBE.
- [Methods, Federated linear regression algorithm (Figure 5)] The claim that fedRBE 'receives results mathematically equivalent' to limma::removeBatchEffect() is asserted but not proven for the missing-value patterns that the tool advertises. limma fits each feature after row-wise deletion of NA samples and uses lm.fit(), i.e., QR decomposition with column pivoting, which can drop dependent columns when the per-feature reduced design is rank deficient. fedRBE instead sums local X^T X and X^T y and solves the normal equations, with masking only for whole batches in which a feature is entirely absent. Partial missingness can make a reduced design rank deficient without removing a whole batch column, for example when a feature is observed in exactly the samples of one batch that have a particular covariate value; in that case the normal equations are singular and the paper specifies no pseudo-inverse or pivoting rule matching limma. Since the simulations contain no missing values and only the proteomics and microarray datasets exercise NAs, the drop-in equivalence claim for arbitrary missingness patterns is not established.
- [Methods, The fedRBE workflow] The stated requirement that covariates not linearly depend on other covariates or on batches applies to the full design matrix, but limma's per-feature fit uses the row-wise reduced design after NA deletion. Even a full-rank complete-data design can become rank deficient after such deletion for a specific feature, so the authors' masking criterion based only on whole-batch feature presence does not cover all supported inputs. The equivalence claim therefore needs either a rigorous proof or a restricted missingness model stated explicitly in the Methods.
minor comments (5)
- [Methods, Proteomics dataset] There is a stray Cyrillic character in 'a total of 98 samplesб'; the manuscript needs a proofreading pass.
- [References] The text cites 'Hardebrodt et al.' for the federated QR decomposition, but the reference list and the published author name use 'Hartebrodt'; please unify the spelling.
- [Methods, Enhancing privacy with additive secret sharing] The description of the secret-sharing protocol says that each client creates c-1 random pieces and then holds 'c pieces in total' after distributing all but one piece; this wording is confusing and should be clarified.
- [Table 1] The table uses '+/—' for d-ComBat's privacy-aware analysis but the legend does not define the symbol; please add a legend or use the same symbols as the other rows.
- [Conclusion] The phrase 'mathematically equivalent' is used where the evidence is numerical equality on tested datasets; unless a proof is supplied, recommend rewording to 'numerically equivalent on the tested datasets'.
Circularity Check
No significant circularity: fedRBE's equivalence claim is benchmarked against the external limma implementation, not defined into existence.
full rationale
The central claim of fedRBE is equivalence to limma::removeBatchEffect() applied to pooled data. This is checked against the original R function on real datasets and simulations; the reported maximum differences (e.g., 3.6E-13) are measurements of an external benchmark, not outputs of a fitted model. The method computes aggregated X^T X and X^T y via secure summation, and the paper explicitly identifies a gap (no federated QR with column pivoting) and then tests empirically that the normal-equation/masking approach matches limma. Thus the equivalence claim is not self-definitional: fedRBE is not defined to match limma by construction, and the validation is against an independent standard. The use of the authors' own prior work (Burankova et al. 2024; Zolotareva et al. 2021) is limited to simulation parameters, dataset provenance, and prior federated-learning context; it does not determine the equivalence result. The concern that partial per-sample missingness could cause rank-deficient reduced designs is a correctness or validation gap, not circularity, because the paper's claim is empirical and falsifiable against limma. No step in the derivation reduces to its own input by construction.
Assumptions & free parameters
assumptions (3)
- ad hoc to paper The mask-based exclusion of whole batches and re-referencing when the reference batch lacks a feature reproduces limma's per-feature row-wise deletion and least-squares fit.
- domain assumption Covariates do not linearly depend on other covariates or on the batches.
- domain assumption Additive secret sharing with at least three clients per feature prevents the coordinator and clients from reconstructing individual client contributions.
Cite this review
Pith. "Pith review of Batch effects can impair federated learning in multi-center omics studies." pith.science (2026). https://pith.science/paper/JK3NEGWO
@misc{pith2026241205894,
author = {Pith},
title = {Pith review of: Batch effects can impair federated learning in multi-center omics studies},
year = {2026},
howpublished = {\url{https://pith.science/paper/JK3NEGWO}},
note = {Machine review of arXiv:2412.05894}
}
read the original abstract
Federated learning (FL) enables collaborative analysis of biomedical data without exchanging sensitive patient-level information, but its performance in multi-center studies may be compromised by batch effects which can obscure biological signals. Here, we systematically assess the impact of uncorrected batch effects on FL outcomes using four multi-center omics datasets, including transcriptomic, proteomic, and metabolomic data, and two representative algorithms: federated k-means clustering and federated random forest classification. Our results demonstrate that uncorrected batch effects undermine unsupervised FL and can substantially degrade supervised FL performance, indicating that privacy-aware batch-effect correction is essential for reliable FL. To enable privacy-preserving BEC in distributed bulk omics data, we introduce fedRBE ( https://featurecloud.ai/app/fedrbe ), a federated implementation of limma's removeBatchEffect() method enhanced by secure multi-party computation, suitable for datasets with missing values and non-identical feature sets across clients, including proteomics and metabolomics data.
Reference graph
Works this paper leans on
-
[1]
Vadapalli, S., Abdelhalim, H., Zeeshan, S. & Ahmed, Z. Artificial intelligence and machine learning approaches using gene expression and variant data for personalized medicine. Brief Bioinform 23 , (2022).2. Vamathevan, J. et al. Applications of machine learning in drug discovery and development. Nat Rev Drug Discov 18 , 463–477 (2019).3. Wang, F., Casali...
-
[41]
Bostami, B. et al. A Decentralized ComBat Algorithm and Applications to Functional Network Connectivity. Front. Neurol. 13 , 826734 (2022).42. Chen, A. A. et al. Privacy-preserving harmonization via distributed ComBat. Neuroimage 248 , 118822 (2022).43. Silva, S., Oxtoby, N., Altmann, A. & Lorenzi, M. Fed-ComBat: A Generalized Federated Framework for Batc...
Reviewed August 11, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.