Multi-Manifolds fusing hyperbolic graph network balanced by pareto optimization for identifying spatial domains of...

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomics enables gene expression analysis within tissue context.
• Graph neural networks are emerging for spatial transcriptomics data.
• Existing methods often overlook complex topological structures in data manifolds.

Purpose of the Study:

• To develop an advanced computational method for identifying spatial domains in spatial transcriptomics data.
• To leverage multi-manifold learning and hyperbolic geometry for enhanced topological feature extraction.
• To improve the accuracy and comprehensiveness of spatial domain identification in gene expression studies.

Main Methods:

• Developed multi-manifold encoders using hyperbolic neural networks.
• Integrated features from distinct manifolds via an attention mechanism.
• Employed Pareto optimization to balance multiple reconstruction losses.

Main Results:

• MManiST consistently outperformed seven state-of-the-art methods on benchmark datasets.
• Ablation experiments validated the effectiveness of individual components and fusion strategies.
• The method successfully identified spatial domains by exploiting deeper topological structures.

Conclusions:

• MManiST offers a superior approach for spatial domain identification in spatial transcriptomics.
• The integration of hyperbolic geometry and multi-manifold learning enhances topological analysis.
• This method provides deeper insights into the pathogenesis of gene expression through improved spatial domain characterization.