Spectral clustering of single-cell multi-omics data on multilayer graphs

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell sequencing generates multimodal data, offering insights into cell heterogeneity.
• Integrating multimodal information for cell type identification presents a significant computational challenge.
• Multilayer graphs naturally represent multi-omic single-cell data, framing cell clustering as a graph partitioning problem.

Purpose of the Study:

• To introduce novel spectral algorithms for clustering cells in multimodal single-cell sequencing datasets.
• To develop a unifying mathematical framework for integrating multiple graph layers in multi-omic data.
• To demonstrate the efficacy of these algorithms and their connection to existing methods.

Main Methods:

• Developed two spectral algorithms: spectral clustering on multilayer graphs and the weighted locally linear (WLL) method.
• Formulated a mathematical framework using Hamiltonian operators and eigenstates to integrate graph layers.
• Reformulated the Seurat weighted nearest neighbor (WNN) algorithm within a multilayer spectral graph theoretic context.

Main Results:

• The WLL method is shown to be a rigorous reformulation of the Seurat WNN algorithm.
• Algorithms were implemented and applied to a CITE-seq dataset of cord blood mononuclear cells.
• Results obtained were comparable to those from the Seurat WNN analysis, validating the new methods.

Conclusions:

• Spectral methods can be effectively extended to multimodal single-cell data analysis.
• The developed algorithms provide a robust approach for cell type identification using multi-omic data.
• This work offers new computational tools for advancing single-cell data interpretation.