Mapping single-cell data to reference atlases by transfer learning.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Large-scale single-cell atlases serve as crucial references for biological studies.
• Challenges in utilizing reference atlases include batch effects, limited computational resources, and data sharing restrictions.

Purpose of the Study:

• To introduce a deep learning strategy, single-cell architectural surgery (scArches), for mapping query datasets onto existing single-cell references.
• To enable efficient, decentralized, and iterative reference building and contextualization of new datasets without raw data sharing.

Main Methods:

• scArches employs transfer learning and parameter optimization for efficient mapping of query datasets.
• The method integrates new datasets onto reference atlases without requiring raw data sharing, addressing privacy and resource limitations.

Main Results:

• scArches effectively removes batch effects while preserving biological state information across diverse atlases (mouse brain, pancreas, immune, whole-organism).
• The approach generalizes to multimodal reference mapping, enabling imputation of missing data modalities.
• scArches successfully retains disease variations (e.g., COVID-19) when mapping to healthy references, facilitating the discovery of disease-specific cell states.

Conclusions:

• scArches provides an efficient and scalable solution for integrating and analyzing single-cell data within large reference atlases.
• This method significantly facilitates collaborative research by enabling iterative construction, updating, and sharing of reference atlases.