AGSI: Adaptive group-enhanced strategy for iterative integration of single-cell multi-omics

Area of Science:

• Computational biology
• Genomics
• Systems biology

Background:

• Single-cell multi-omics data integration is crucial for understanding cellular heterogeneity and disease mechanisms.
• Existing methods struggle with uniform gene correspondence evaluation and static integration, failing to capture cell-level heterogeneity.
• Biological systems exhibit modular organization and varying cell-level heterogeneity, necessitating advanced integration approaches.

Purpose of the Study:

• To propose AGSI, an adaptive framework for robust multi-omics integration.
• To address limitations in current methods by focusing on co-regulated gene modules and iterative reliability assessment.
• To improve the accuracy and interpretability of single-cell multi-omics data integration.

Main Methods:

• AGSI utilizes Latent Dirichlet Allocation to identify co-regulated gene modules.
• Cross-modal correspondence is evaluated at the module level using Wasserstein-enhanced similarity metrics and dual reliability modeling.
• Adaptive thresholding dynamically refines cell selection and integration throughout an iterative process.

Main Results:

• AGSI significantly outperforms seven state-of-the-art methods across multiple datasets (PBMC, mouse brain, human myocardial infarction).
• Achieved up to 25.6% F1 improvement over baseline and 11.9% over the best competing method on neural datasets.
• Demonstrated robustness, maintaining over 85% accuracy even with 50% data dropout.

Conclusions:

• AGSI offers a robust, scalable, and interpretable solution for multi-omics integration.
• The framework achieves superior technical performance, suitable for biomedical analyses.
• AGSI is particularly well-suited for accurate cell type identification in complex biological systems.