scHG: A supercell framework with high-order graph learning enables scalable multi-omics analysis.

Area of Science:

• Computational biology
• Single-cell genomics
• Multi-omics data integration

Background:

• Multi-omics profiling offers detailed cellular insights but faces challenges in identifying rare cell populations during large-scale clustering.
• Existing methods struggle to maintain computational tractability while preserving subtle biological structures across diverse omics data.

Purpose of the Study:

• To develop a novel computational framework for enhanced multi-omics clustering and rare cell population identification.
• To introduce the supercell paradigm for grouping expression-coherent cells into intermediate units.

Main Methods:

• The supercell paradigm groups cells into intermediate units using angle-aware similarity and co-occurrence metrics.
• The scHG framework employs a high-order graph learning approach with an omics-weighted optimizer for adaptive data integration.
• Scalability is achieved through sparse matrix optimization and iterative graph refinement.

Main Results:

• scHG consistently outperforms state-of-the-art methods across six benchmark datasets, improving clustering accuracy (ARI, NMI) and reducing runtime.
• The method successfully resolves fine-grained heterogeneity within T-cell populations.
• scHG uncovers rare cell populations, such as dendritic cells and NK-like B cells, previously hidden by standard clustering.

Conclusions:

• Supercells offer an efficient intermediate representation for large-scale multi-omics integration.
• The scHG framework provides a practical and scalable mechanism for sensitive rare-cell detection in complex single-cell datasets.