GatorDuo: Global-Consistency Dual-Graph Refinement With Pseudo-Label Agreement for Spatial Transcriptomics.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomics (ST) integrates gene expression with spatial data for tissue domain identification.
• Current graph-based methods struggle with noisy ST data, leading to inaccurate neighborhood construction and blurred domain boundaries.
• Spurious edges in static graphs propagate misleading signals, weakening cluster separability.

Purpose of the Study:

• To introduce GatorDuo, a novel framework for robust spatial domain identification in ST data.
• To enhance the accuracy of neighborhood construction and stabilize graph topology for improved representation learning.
• To adaptively determine optimal clustering granularity and generate unified embeddings for downstream analysis.

Main Methods:

• GatorDuo employs a topology-aware dual-graph contrastive self-supervised framework.
• It utilizes global-consistency-based graph refinement with a pseudo-label agreement mask to suppress spurious edges.
• A contextual bandit reinforcement-learning strategy adaptively selects clustering granularity, and a Mixture-of-Experts module integrates embeddings.

Main Results:

• GatorDuo consistently achieves strong and robust spatial domain identification performance across eight diverse ST benchmarks.
• The method outperforms ten representative baselines in multiple clustering metrics.
• It successfully suppresses cross-domain shortcut edges, stabilizing neighborhood topology for reliable representation learning.

Conclusions:

• GatorDuo offers a robust and accurate solution for spatial domain identification in spatial transcriptomics.
• The framework's adaptive clustering and unified embedding generation support comprehensive downstream biological analyses.
• This approach enhances the reliability of ST data interpretation, particularly in the presence of sparsity and noise.