scArches

scArches integrates newly produced single-cell RNA sequencing datasets into existing reference atlases using transfer learning to remove batch effects while preserving biological variation.

Key Features:

• Integration of Single-Cell Data: scArches maps new single-cell RNA sequencing datasets onto existing reference atlases.
• Compatibility with scanpy: scArches is compatible with scanpy for single-cell RNA-seq data analysis workflows.
• Conditional Generative Models: scArches provides efficient implementations of conditional generative models tailored to single-cell data.
• Transfer Learning and Parameter Optimization: scArches leverages transfer learning and parameter optimization to enable parameter-efficient mapping of query datasets.
• Decentralized and Iterative Reference Building: scArches supports decentralized, iterative construction and updating of reference atlases without sharing raw data.
• Batch Effect Removal: scArches removes batch effects while preserving nuanced biological state information, using fewer parameters than de novo integration methods.
• Preservation of Biological Variation: scArches preserves disease-specific variation such as COVID-19-related signals and can reveal novel cell identities not present during training.
• Mapping Query Datasets: scArches maps query datasets onto reference atlases to maintain high fidelity of biological information.

Scientific Applications:

• Reference Atlas Construction and Updating: scArches supports iterative construction and updating of reference atlases for collaborative research projects.
• Contextualization of New Datasets: scArches enables contextualization of smaller-scale studies by mapping them to large single-cell reference atlases.
• Disease Variation Mapping: scArches preserves and maps disease-specific variations, exemplified by COVID-19 disease variation mapping, aiding discovery of novel cell identities and disease-associated states.
• Large-Scale Collaborative Integration: scArches facilitates integration of datasets from multiple contributing groups in large collaborative projects without raw data sharing.

Methodology:

scArches applies deep learning, including conditional generative models, and leverages transfer learning and parameter optimization to map query single-cell RNA-seq datasets onto reference atlases, supporting decentralized training and mitigating batch effects.