scUCAF: An uncertainty-aware cross-omics alignment and fusion network for single-cell multi-omics data clustering

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell multi-omics sequencing reveals cellular heterogeneity.
• Cell clustering is vital for multi-omics data analysis.
• Existing methods struggle with varying data quality across omics, impacting feature representation.

Purpose of the Study:

• To propose scUCAF, an uncertainty-aware network for robust multi-omics clustering.
• To address noise and data quality variations in single-cell multi-omics data.
• To generate reliable cell representations for improved downstream analyses.

Main Methods:

• Utilized a variational autoencoder with negative binomial distribution for noise-robust feature extraction.
• Implemented high-confidence cluster-guided contrastive learning for cross-omics consistency.
• Developed an uncertainty-aware fusion and gating network for dynamic, bias-mitigated omics integration.

Main Results:

• scUCAF demonstrated superior performance on eight single-cell multi-omics datasets compared to existing methods.
• The method effectively mitigates biases from low-quality data.
• Clustering results were validated through downstream cell type annotation and biomarker identification.

Conclusions:

• scUCAF provides reliable cell representations for multi-omics clustering.
• The uncertainty-aware approach enhances robustness against data quality variations.
• This method advances cell type annotation and biomarker discovery in complex biological systems like liver cancer.