Benchmarking component choices for unpaired single cell RNA and epigenomic integration

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Single-cell multi-omics technologies offer deep biological insights but integrating unpaired data remains challenging.
• Existing computational methods for unpaired data integration lack systematic evaluation.
• Paired multi-omics datasets now enable rigorous benchmarking of unpaired integration pipelines.

Purpose of the Study:

• To systematically evaluate computational strategies for integrating unpaired single-cell RNA sequencing (scRNA-seq) with peak-based epigenomic (scATAC-seq, ChIP-seq) data.
• To establish a robust and generalizable computational pipeline for unpaired multi-omics integration.
• To identify optimal choices for feature linking, dimension reduction, and clustering in unpaired data integration.

Main Methods:

• Leveraged paired scRNA-seq, scATAC-seq, and ChIP-seq datasets for systematic evaluation.
• Benchmarked individual integration steps (feature linking, dimension reduction, clustering) and their combinations.
• Assessed performance using metrics relevant to preserving cellular neighborhoods and molecular dynamics.

Main Results:

• Gene activity scores showed limited correlation with gene expression but effectively preserved cellular neighborhoods for clustering.
• Dimension reduction was the most critical step, with non-linear methods generally outperforming linear methods.
• Optimal transport (OT)-based label transfer consistently outperformed other strategies across various embeddings.

Conclusions:

• This benchmark provides crucial insights for developing robust methods for unpaired multi-omics data integration.
• Findings guide the selection of optimal computational strategies for complex multi-omics studies.
• The study facilitates a more comprehensive understanding of cell states and molecular dynamics from unpaired data.