scMAG: Integrating single-cell multi-omics data via multi-stage deep fusion with manifold-aware gating

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell omics technologies allow simultaneous profiling of multiple genomic modalities.
• Integrating multi-omics data requires computational frameworks balancing cross-modal associations and biological fidelity.
• A key challenge is harmonizing omics layer alignment with modality-specific distribution preservation.

Purpose of the Study:

• To propose an innovative deep fusion method for single-cell multi-omics data integration.
• To develop the scMAG algorithm for improved clustering accuracy and data visualization.
• To achieve adaptive alignment of multi-omics latent spaces and optimize data distribution, suppressing noise.

Main Methods:

• A multi-stage feature deep fusion formalization method.
• Omics multi-core manifold preservation strategy.
• Guided gating optimization strategy for adaptive alignment and distribution optimization.

Main Results:

• scMAG demonstrated superior performance in clustering and data visualization compared to benchmark algorithms using scRNA-seq, scATAT-seq, and ADT datasets.
• The algorithm effectively balances shared and modality-specific signals across multi-omics data.
• Enhanced feature distribution in the data space contributed to improved performance.

Conclusions:

• scMAG offers significant improvements in clustering, dimensionality reduction, batch effect removal, and multimodal data integration.
• The algorithm exhibits good biological interpretability, providing theoretical support for multimodal single-cell data analysis.
• scMAG effectively suppresses biological noise and measurement errors in single-cell multi-omics data.