Understanding gene activity at the single-cell level is a powerful tool in modern biology, but comparing data across different studies can be challenging. One major issue is “batch effects,” which are technical differences introduced when samples are processed or sequenced at different times or in different labs. These variations can make it hard to tell whether observed changes are due to real biological differences or just experimental inconsistencies.

Researchers from the University of Hamburg have developed a new tool called FedscGen to address this problem. FedscGen helps scientists correct for batch effects in single-cell RNA sequencing data, allowing results from different experiments to be compared more accurately. What makes FedscGen unique is that it uses federated learning, a privacy-protecting method where data can be analyzed collaboratively without needing to be centralized or directly shared. This is especially important when working with sensitive clinical data.

FedscGen delivers results comparable to scGen without compromising patient privacy

a FedscGen orchestrates federated training by deploying models to clients for local updates, followed by secure aggregation over communication rounds. The correction workflow allows FedscGen to correct batch effects and support the inclusion of new studies. b FedscGen achieves comparable batch effect correction to scGen across batch mixing and biological preservation metrics for different held-out batches in the Human Pancreas (HP) dataset. In the held-out scenario, the training WF includes four clients (each with a batch), and the correction WF incorporates the fifth batch as a new study for correction. In the no held-out scenario, all five batches are included as clients in both the training and correction workflows. c UMAP visualization of batch effect correction using FedscGen in a federated scenario with five clients, showing effective batch mixing and cell type separation. d UMAP plots for each held-out batch scenario demonstrating that FedscGen achieves correction quality comparable to centralized scGen without a held-out batch

FedscGen builds on an existing model called scGen and improves it by adding secure multiparty computation, which ensures that even during collaboration, individual datasets remain private. The team tested the tool using a well-known Human Pancreas dataset and found that FedscGen performs just as well as existing methods using multiple accuracy and consistency metrics.

Best of all, FedscGen is available as a ready-to-use app through the FeatureCloud platform, making it easier for researchers around the world to collaborate securely on RNA sequencing projects.

Availability – The source code for FedscGen is available at https://github.com/Mohammad-Bakhtiari/FedscGen