FedscGen: privacy-aware federated batch effect correction of single-cell RNA sequencing data

Abstract scRNA-seq data from clinical samples are prone to batch effects, while hospitals are hesitant to share their data for centralized analysis, including batch effect correction, due to the privacy sensitivity of human genomic data. We present FedscGen, a novel privacy-aware federated method based on the generative integration approach scGen. FedscGen presents two federated workflows for training and correction of batch effects with inclusion of new studies. We benchmark FedscGen and scGen using eight datasets and nine metrics to demonstrate competitive results. On the Human Pancreas dataset, for instance, the performance difference of all models is zero for NMI, GC, ILF1, ASW_C, and kBET while FedscGen outperforms by 0.03 in EBM. FedscGen opens a privacy-preserving path for single-cell RNAseq batch effect correction in particular in clinical multi-center studies. FedscGen is published as a FeatureCloud app to be used in real world federated collaboration (https://featurecloud.ai/app/fedscgen).