Network methods for diagonal integration of unpaired single-cell multiomics data: a review

Area of Science:

• Computational Biology
• Genomics
• Proteomics

Background:

• Single-cell sequencing enables multi-omics, but mass spectrometry-based single-cell proteomics (scMS) is destructive, preventing simultaneous transcriptomic capture.
• scMS offers unbiased, genome-scale proteome coverage, unlike targeted antibody-based methods, but requires post hoc integration of unpaired datasets.
• The diagonal integration of unpaired single-cell transcriptomics and proteomics data is a critical challenge not well-covered by existing reviews.

Purpose of the Study:

• To survey the computational pipeline for constructing proteogenomic networks from unpaired single-cell data.
• To provide a comprehensive overview of unimodal and cross-modal integration strategies.
• To outline benchmarking paradigms and identify future development directions in single-cell proteogenomics.

Main Methods:

• Survey of unimodal network inference techniques (knowledge-based, probabilistic graphical models, generative models).
• Review of cross-modal integration architectures (network propagation, graph neural networks, consensus frameworks).
• Analysis of benchmarking paradigms for network reconstruction and multi-task integration.

Main Results:

• Detailed survey of computational pipelines for proteogenomic network construction from unpaired single-cell data.
• Identification of key methods for unimodal and cross-modal integration, including specific algorithms like scMRDR and scmFormer.
• Evaluation of benchmarking tools such as BEELINE, GRETA, and scMultiBench for assessing integration performance.

Conclusions:

• Three key areas for future development are identified: generative proteomic translation, inductive prior embedding, and perturbation-based causal benchmarking.
• The review provides a foundation for researchers tackling the challenge of integrating unpaired single-cell omics data.
• Guidance on metric selection for network sparsity and class imbalance in benchmarking is offered.