Technical Summary: Batch Effects Impair Federated Learning in Multi-Center Omics Studies

Problem Statement

Federated learning (FL) offers a privacy-preserving framework for collaborative biomedical analysis, allowing institutions to train models without exchanging sensitive patient-level data. However, the performance of FL in multi-center omics studies is significantly compromised by batch effects—systematic biases arising from differences in experimental conditions, equipment, or protocols across sites.

While batch effect correction (BEC) is standard in centralized analyses, its application in decentralized settings remains under-characterized. Existing decentralized BEC tools face critical limitations:

• Applicability: Methods like FedscGen and federated Harmony are designed for single-cell RNA-seq and may not suit bulk omics data, which often has smaller sample sizes.
• Security: Tools like d-ComBat require manual data coordination and expose design matrices, creating vulnerabilities to reconstruction attacks.
• Data Handling: Most existing tools require preprocessing or imputation of missing values, rendering them unsuitable for bulk omics datasets (e.g., proteomics and metabolomics) that frequently contain extensive non-random missingness.

The authors hypothesize that uncorrected batch effects in decentralized data adversely affect downstream federated analyses, necessitating a privacy-aware, robust BEC solution.

Methodology: The fedRBE Framework

To address these challenges, the authors introduce fedRBE, a federated implementation of the removeBatchEffect() method from the popular limma R package. The framework integrates Federated Learning (FL) with Secure Multi-Party Computation (SMPC) to ensure data privacy while enabling accurate batch correction.

Core Technical Components

1. Algorithmic Foundation:

   • fedRBE utilizes linear modeling to adjust for batch effects and known biological covariates.
   • Unlike ComBat-based approaches, it natively supports datasets with missing values without requiring prior imputation, a critical feature for proteomics and metabolomics.
   • It handles non-identical feature sets across clients (e.g., different genes or proteins detected at different sites) by computing the union of features for correction.

2. Privacy-Preserving Workflow (FL + SMPC):

   • Setup: A coordinator initiates the workflow via the FeatureCloud platform. Participants upload local data (.csv) and configuration files (.yml) to a local Docker container; raw data never leaves the site.
   • Feature Selection: The coordinator identifies the intersection of covariates and the union of features available across at least three clients (a requirement for the SMPC protocol).
   • Secure Aggregation:
     • Each client computes local intermediate matrices ($X^T X$ and $X^T y$) based on their local data and design matrix.
     • Additive Secret Sharing: To prevent reconstruction attacks, clients split these matrices into random shares. These shares are exchanged and aggregated such that the coordinator receives only the global sums ($\sum X^T X$ and $\sum X^T y$) without access to individual client contributions.
   • Correction: The coordinator calculates global regression coefficients ($\beta$) from the aggregated sums and broadcasts them back to clients. Each client then applies these coefficients to correct their local data.

3. Validation Strategy:

   • The authors evaluated fedRBE on four real-world multi-center datasets:
     1. MS-based proteomics (E. coli).
     2. Microarray-based ovarian cancer transcriptomics.
     3. ccRCC proteomics.
     4. Quartet multi-omics (transcriptomics, proteomics, metabolomics).
   • They compared fedRBE outputs against centralized limma results and assessed downstream performance using federated k-means clustering (unsupervised) and federated random forest classification (supervised).
   • Two training scenarios were tested: within-client splits (70:30) and leave-one-center-out (training on $N-1$ centers, testing on the held-out center).

Key Results

1. Numerical Equivalence to Centralized BEC

fedRBE produces results mathematically equivalent to applying limma::removeBatchEffect() to pooled data.

• Precision: The maximum absolute element-wise difference between fedRBE and centralized results was negligible ($\le 2.49 \times 10^{-13}$).
• Robustness: This equivalence held true even under simulated scenarios with strong data imbalance across centers.
• Variance Reduction: After fedRBE correction, batch-specific variance dropped from 15–99% to essentially zero across all datasets.

2. Impact on Federated Learning Performance

The study demonstrates that uncorrected batch effects severely degrade FL performance, particularly in scenarios where models must generalize across unseen centers.

• Unsupervised Learning (k-means):

  • Without BEC, clustering was dominated by technical batch differences rather than biological signals, failing to identify target classes.
  • With fedRBE, the Adjusted Rand Index (ARI) improved significantly, reaching median values of 0.7–1.0, successfully recovering biologically meaningful structures.

• Supervised Learning (Random Forest):

  • Leave-One-Center-Out: BEC was critical for generalization. In the E. coli dataset, applying BEC increased the median Matthews Correlation Coefficient (MCC) from 0 (random prediction) to 1 (perfect classification). It also reduced result variability and stabilized model convergence.
  • Within-Client Split: Performance was high (near perfect) in both corrected and uncorrected scenarios, suggesting that models can learn batch-specific patterns when the test set shares the same batch characteristics as the training set. However, this does not translate to generalizability across centers.

Significance and Claims

The paper claims that privacy-aware batch effect correction is essential for reliable federated learning in multi-center omics studies. The specific contributions and significance include:

1. First Automated Tool for Bulk Omics: fedRBE is presented as the first fully automated BEC tool for bulk omics data that supports missing values and non-identical feature sets across clients while maintaining privacy.
2. Privacy Preservation: By combining FL with SMPC, fedRBE prevents reconstruction attacks and ensures that no raw data or individual design matrices are exposed, addressing security limitations of previous tools like d-ComBat.
3. Practical Usability: Implemented as a FeatureCloud application with a graphical user interface (GUI), fedRBE lowers the barrier to entry for researchers with diverse computational backgrounds, enabling decentralized workflows that were previously infeasible due to batch effects.
4. Empirical Validation: The study provides the first systematic assessment showing that uncorrected batch effects can render unsupervised FL useless and substantially degrade supervised FL generalization, validating the necessity of tools like fedRBE.

The authors maintain a modest tone regarding limitations, acknowledging that fedRBE inherits the linear modeling constraints of limma (potentially insufficient for non-linear batch effects or count-based RNA-seq) and that BEC in unsupervised settings requires caution to avoid over-correction of genuine biological variation.