FedRBE -- a decentralized privacy-preserving federated batch effect correction tool for omics data based on limma

TL;DR

fedRBE is a federated, privacy-preserving tool for batch effect correction in omics data, supporting missing values and automated workflows, enabling collaborative analysis without data sharing.

Contribution

It introduces fedRBE, a novel federated implementation of limma's batch correction method that handles missing data and enhances privacy through secure multi-party computation.

Findings

Performance comparable to centralized methods with negligible differences

Effectively handles missing values in omics data

Facilitates large-scale collaborative studies without data sharing

Abstract

Batch effects in omics data obscure true biological signals and constitute a major challenge for privacy-preserving analyses of distributed patient data. Existing batch effect correction methods either require data centralization, which may easily conflict with privacy requirements, or lack support for missing values and automated workflows. To bridge this gap, we developed fedRBE, a federated implementation of limma's removeBatchEffect method. We implemented it as an app for the FeatureCloud platform. Unlike its existing analogs, fedRBE effectively handles data with missing values and offers an automated, user-friendly online user interface (https://featurecloud.ai/app/fedrbe). Leveraging secure multi-party computation provides enhanced security guarantees over classical federated learning approaches. We evaluated our fedRBE algorithm on simulated and real omics data, achieving performance comparable to the centralized method with negligible differences (no greater than 3.6E-13). By enabling collaborative correction without data sharing, fedRBE facilitates large-scale omics studies where batch effect correction is crucial.