scJoint integrates atlas-scale single-cell RNA-seq and ATAC-seq data with transfer learning

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Single-cell multiomics data is rapidly expanding, presenting integration challenges for large, heterogeneous cell atlases.
• Existing methods struggle with the complexity and scale of integrating diverse data modalities from the same tissue.

Purpose of the Study:

• To develop a computational method for integrating atlas-scale, heterogeneous single-cell RNA sequencing (scRNA-seq) and single-cell assay for transposase-accessible chromatin sequencing (scATAC-seq) data.
• To leverage transfer learning for improved cell-type annotation and joint visualization of multiomic single-cell data.

Main Methods:

• Introduced scJoint, a transfer learning approach utilizing a semisupervised framework and neural networks.
• scJoint simultaneously trains labeled and unlabeled data, enabling label transfer from annotated scRNA-seq to unlabeled scATAC-seq data.
• The method was validated on atlas-scale data and multimodal datasets (ASAP-seq, CITE-seq).

Main Results:

• scJoint demonstrated computational efficiency in integrating heterogeneous single-cell multiomic data.
• Achieved substantially higher cell-type label accuracy compared to existing methods.
• Provided meaningful joint visualizations, overcoming data modality heterogeneity.

Conclusions:

• scJoint effectively integrates large-scale, heterogeneous scRNA-seq and scATAC-seq data.
• The method enhances cell-type identification accuracy and facilitates comprehensive understanding of cellular phenotypes through joint visualization.
• scJoint represents a significant advancement in analyzing complex single-cell multiomic datasets.