Online single-cell data integration through projecting heterogeneous datasets into a common cell-embedding space.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Single-cell experiments generate large, heterogeneous datasets requiring advanced integration methods.
• Existing computational tools struggle with increasing data scale and the need for continuous updates.
• Cross-referencing new data with foundational datasets is crucial for building comprehensive atlases.

Purpose of the Study:

• To develop a deep-learning method for online integration of diverse single-cell data.
• To create a batch-invariant, common cell-embedding space for seamless data integration.
• To enable the continuous expansion of single-cell atlases with new data.

Main Methods:

• SCALEX, a deep-learning approach, projects single-cell data into a common embedding space.
• The method operates in a truly online manner, without requiring model retraining.
• Evaluated on diverse single-cell modalities including single-cell RNA sequencing (scRNA-seq) and single-cell assay for transposase-accessible chromatin sequencing (scATAC-seq).

Main Results:

• SCALEX significantly outperforms online non-negative matrix factorization (iNMF) and other state-of-the-art integration methods.
• Demonstrates superior performance on datasets with partial overlaps, accurately aligning cell populations.
• Successfully used to construct continuously expandable single-cell atlases for human, mouse, and COVID-19 patients.

Conclusions:

• SCALEX offers a powerful solution for large-scale single-cell data integration.
• Its online integration capability is ideal for building and expanding dynamic single-cell atlases.
• Facilitates building upon previous scientific insights by enabling continuous data incorporation.