Curriculum-guided divergence scheduling improves single-cell clustering robustness

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Single-cell RNA sequencing (scRNA-seq) data presents significant challenges for deep clustering due to extreme sparsity and noise.
• Existing static deep clustering methods struggle to effectively handle the complexities of scRNA-seq data.

Purpose of the Study:

• To introduce DAGCL (Dynamic Attention-enhanced Graph Embedding with Curriculum Learning), a novel framework for robust deep clustering of scRNA-seq data.
• To address limitations of static paradigms by reframing representation learning as an evolutionary process.

Main Methods:

• DAGCL utilizes a dynamic graph embedding approach with a curriculum-guided scheduling mechanism.
• It actively modulates attention intensity and supervision stringency during training to align model complexity with feature maturity.
• Incorporates an entropy-regularized Sinkhorn projection for globally balanced soft assignments and stable optimization.

Main Results:

• DAGCL consistently outperforms existing baseline methods across 27 benchmark datasets.
• Demonstrates superior clustering accuracy and robustness in analyzing scRNA-seq data.
• Effectively mitigates early-stage confirmation bias through its dynamic training strategy.

Conclusions:

• DAGCL establishes a principled strategy for unsupervised learning in bioinformatics.
• The framework enables co-evolution of structural constraints and supervisory pressure with learned representations.
• Offers a robust solution for deep clustering of challenging sparse and noisy single-cell data.