Single-cell multi-omics integration for unpaired data by a siamese network with graph-based contrastive loss.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell omics technologies enable epigenome, genome, and transcriptome measurement.
• Integrating multi-omics data at single-cell resolution presents significant challenges.

Purpose of the Study:

• To develop a novel deep learning framework for single-cell multi-omics data integration.
• To address challenges in combining diverse single-cell omics datasets.

Main Methods:

• Proposed MinNet, a Siamese neural network variation.
• Utilized graph-based contrastive loss for training.
• Applied to integrate single-cell RNA sequencing (scRNA-seq) with single-cell Assay for Transposase-Accessible Chromatin sequencing (scATAC-seq) and epitope data.

Main Results:

• Demonstrated accuracy and generalizability in integrating scRNA-seq with scATAC-seq and epitope data.
• Showcased effective removal of batch effects.
• Developed downstream analysis methods (smoothing, cis-regulatory element inference) validated with external data.
• Applied to a COVID-19 dataset, highlighting the necessity of integration-based analysis.

Conclusions:

• MinNet is a top-performing deep learning framework for single-cell multi-omics integration.
• Effective for datasets with batch and biological variances.
• Facilitates analysis of genome-transcriptome interplay at single-cell resolution.