💻 bioinformatics

FedEdgeR: federated and privacy-preserving edgeR for differential gene expression analysis

The paper introduces FedEdgeR, a secure multi-party computation-based federated learning framework that enables privacy-preserving differential gene expression analysis using edgeR, demonstrating performance equivalent to centralized analysis and superior to traditional meta-analysis methods across diverse RNA-seq datasets.

Original authors: Song, X., Wei, Z.

Published 2026-09-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Song, X., Wei, Z.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

In the world of modern medicine, understanding how genes behave is like reading the instruction manual for a living cell. When scientists want to find out which genes are turned on or off during a disease, they use a technique called RNA sequencing to count the molecular messages inside cells. To get a clear picture, researchers often need to combine data from many different hospitals or labs, creating a massive pool of information that reveals patterns too subtle to see in a single study. However, a strict barrier stands in the way: patient privacy. Laws and ethical rules prevent hospitals from sharing the raw genetic data of their patients, as this information could potentially be used to identify individuals. This creates a difficult dilemma: how can scientists gain the power of a huge combined study without ever seeing the private data of a single patient?

For years, the standard solution was a method called meta-analysis. In this approach, each hospital runs its own analysis separately and sends only the final results, such as a list of important genes, to a central team. The central team then tries to stitch these separate lists together. While this protects privacy, it often weakens the science. If one hospital has very few patients or an unbalanced mix of healthy and sick individuals, the local analysis might fail to find the truth, and the central team cannot fix those missing pieces. It is like trying to solve a giant puzzle by only looking at the corner pieces from different boxes; you might get the edges, but the middle remains blurry.

A new approach called federated learning offers a different path. Instead of sending results back and forth, this method allows computers at different locations to work together on the same mathematical problem without ever moving the raw data. They share only the intermediate steps of their calculations, which are scrambled to hide the original numbers, and then combine these scrambled pieces to get a final answer. This is mathematically equivalent to having all the data in one place, but the data never leaves its home. While this technique has been applied to some gene analysis tools, a major and widely used method called edgeR, which is particularly good for studies with small numbers of patients or highly variable biological samples, had no such federated version. This left a significant gap in the ability to study difficult or rare conditions across multiple centers.

To fill this gap, researchers at the New Jersey Institute of Technology developed a new system called FedEdgeR. This tool brings the power of the edgeR method into a secure, federated environment. The team faced a unique challenge because the edgeR method does not just calculate a result in a single pass; it uses a complex, step-by-step process that repeatedly refines its estimates to find the most accurate answer. This iterative nature made it much harder to split the work across different computers without leaking private information. The researchers solved this by designing a system where each hospital's computer performs its local calculations, scrambles the results with random noise, and sends them to a central coordinator. A separate server handles the noise, allowing the coordinator to remove it and reveal the true combined result without ever seeing the individual numbers from any single hospital.

The team tested this new system on four real-world datasets involving thousands of genes and patients from various studies, including research on breast cancer, melanoma, and liver disease. They compared the results of FedEdgeR against the "gold standard" of pooling all the raw data together in one place, which is what scientists would do if privacy laws did not exist. The results were strikingly precise. In every test, the federated system produced results that were virtually identical to the pooled analysis. The lists of important genes matched perfectly, and the statistical confidence in those findings was the same. Even in a very difficult test case with only six patients split across three locations, where traditional methods would have failed completely because there was not enough data at any single site, the new system succeeded. It reconstructed the full picture by combining the small pieces of information from each site before the analysis began, rather than trying to combine the final answers afterward.

When the researchers compared FedEdgeR to the older meta-analysis methods, the difference in performance was clear. The traditional methods, which combine final lists of genes, often missed important signals or produced confusing rankings, especially when the data from different sites was uneven. In contrast, the new federated system consistently outperformed these older techniques, recovering the same high-quality results as if all the data had been merged from the start. The system proved that it is possible to conduct powerful, large-scale genetic studies across many institutions without compromising patient privacy. By enabling scientists to use the most robust statistical tools available without needing to move sensitive data, this work removes a major barrier to collaborative medical research, allowing for deeper insights into disease mechanisms that were previously out of reach.

Problem
Differential expression (DE) analysis of RNA-seq data is fundamental to characterizing transcriptomic responses, with edgeR being a preferred framework for datasets with small sample sizes or high biological variability due to its negative binomial generalized linear model (GLM) and empirical Bayes shrinkage. However, multi-center studies are often hindered by privacy regulations (e.g., GDPR) that prevent the sharing of patient-level raw data. While meta-analysis methods (e.g., Fisher's, Stouffer's, random-effects models) avoid raw data sharing, they treat sites as independent studies, leading to a loss of statistical power, particularly when sample sizes are imbalanced or sites have biased case-to-control ratios. Existing federated learning solutions for RNA-seq, such as Flimma (for limma-voom) and FedPyDESeq2 (for DESeq2), do not cover edgeR. Federating edgeR is uniquely challenging because it relies on iteratively reweighted least squares (IRLS) for GLM fitting and Cox-Reid adjusted profile likelihood (APL) for dispersion estimation, requiring multiple rounds of communication and complex non-linear optimization, unlike the single-pass weighted least squares used in limma-voom.

Methodology
The authors present FedEdgeR, the first federated implementation of the edgeR DE pipeline, protected by Secure Multi-Party Computation (SMPC). The system utilizes a hybrid federated architecture involving three parties: $K$ client sites, an aggregator, and a compensator.

Privacy Model: Clients hold local count and design matrices. To share statistics without revealing raw data, clients mask local statistics ( $M_i$ ) with random noise ( $N_i$ ) before sending $M_i + N_i$ to the aggregator and $N_i$ to the compensator. The aggregator recovers the global sum $\sum M_i$ by subtracting the global noise sum provided by the compensator. The threat model assumes all parties are "honest-but-curious" and do not collude.
Federated Algorithm: The pipeline mirrors the centralized edgeR workflow:
1. Normalization: Uses offsets (e.g., from TMM normalization) as input.
2. GLM Fitting (IRLS): The negative binomial GLM is fitted via IRLS. The authors demonstrate that the sufficient statistics required for the IRLS update ( $X^T W X$ and $X^T W z$ ) are additive across sites. Clients compute local terms based on the current coefficients broadcast by the aggregator; these are securely aggregated to update the global coefficients iteratively.
3. Dispersion Estimation: Supports common, trended, and tagwise (gene-specific) dispersions. The Cox-Reid APL is computed by aggregating local log-likelihoods and Fisher information matrices. Trended dispersions are derived via local-regression smoothing on the federated data, and tagwise dispersions are obtained via empirical Bayes shrinkage.
4. Likelihood Ratio Test: Full and null models are fitted, and the likelihood ratio statistic is computed from aggregated residual deviances.
Implementation: The core federated arithmetic and SMPC masking are implemented in Python (NumPy/SciPy), interfacing with the original R edgeR package for specific routines like trended dispersion smoothing and empirical Bayes shrinkage.

Key Contributions

First Federated edgeR: It is the first DE pipeline to combine SMPC privacy protection with the non-linear, iterative model fitting required by edgeR (IRLS and Cox-Reid APL).
Mathematical Derivation: The authors provide a complete derivation showing that the sufficient statistics for IRLS and APL are additive across sites, enabling secure aggregation without loss of information.
Performance Validation: The method is evaluated on four representative RNA-seq datasets (BRCA, Mel-ICI, GSE144269, and a small paired Oral dataset) against pooled edgeR (the gold standard) and four meta-analysis baselines.

Results

Equivalence to Pooled Analysis: FedEdgeR achieves near-machine precision equivalence with centralized pooled edgeR across all datasets.
- Pearson correlation for $-\log_{10}(p\text{-values})$ is $\ge 0.99999$ (often 1.00000000).
- Top-100 differentially expressed (DE) gene overlap is 100%.
- F1 score at nominal FDR < 0.05 is 1.000.
- Dispersion estimates (common, trended, and tagwise) match the centralized pipeline with negligible error (e.g., RMSE $< 10^{-7}$ ).
Superiority over Meta-Analysis: FedEdgeR consistently outperforms Fisher's method, Stouffer's method, the random-effects model, and RankProd.
- Meta-analysis methods suffer significant performance drops, particularly in small-sample or imbalanced scenarios (e.g., the 6-sample Oral dataset where per-site models are saturated).
- In the Oral dataset, meta-analysis F1 scores ranged from 0.067 to 0.878, whereas FedEdgeR maintained an F1 of 1.000.
Robustness: Accuracy remains stable across varying numbers of client sites ( $K=2$ to $10$) and under size- or ratio-imbalance partitions.
Efficiency: The end-to-end runtime is dominated by the common-dispersion grid search, ranging from ~21 seconds (6 samples) to ~8.6 minutes (227 samples). Communication overhead is minimal (under 2.5 MB per round per client).

Significance
The paper claims that FedEdgeR closes the federation gap for edgeR, the last major DE framework lacking a decentralized counterpart. By enabling mathematically equivalent results to pooled analysis without sharing raw data, FedEdgeR allows multi-center studies to retain full statistical power, even in challenging scenarios with small sample sizes, high variability, or severe data imbalance where meta-analysis fails. The work demonstrates that complex, iterative statistical procedures like IRLS and Cox-Reid APL can be securely federated, offering a privacy-preserving alternative that avoids the information loss inherent in traditional meta-analysis approaches.

More like this