The paper introduces scArches, a novel deep learning strategy based on transfer learning (TL) and architectural surgery, designed for efficiently mapping new, or query, single-cell datasets onto existing, large-scale reference atlases. This method addresses the challenges in single-cell genomics of batch effects between datasets, limited computational resources, and data sharing restrictions by only requiring fine-tuning a small subset of the network parameters (adaptors). The scArches pipeline enables decentralized, iterative building and updating of reference models without needing to share raw data, offering significant speed and efficiency advantages over traditional, full data integration techniques. Furthermore, the paper demonstrates scArches' capability to preserve biological variation—including disease-specific cell states and rare cell types—while effectively removing technical batch effects, and facilitates knowledge transfer for cell type annotation and imputing missing data modalities.

References:

• Lotfollahi M, Naghipourfar M, Luecken M D, et al. Mapping single-cell data to reference atlases by transfer learning[J]. Nature biotechnology, 2022, 40(1): 121-130.