DWGCN: distance-weighted graph convolutional network for robust spatial domain identification in spatial...

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Graph Convolutional Networks (GCNs) are key for spatial transcriptomics (ST) but struggle with uniform neighbor weighting.
• Standard GCN methods suppress spatial heterogeneity, reduce resolution, and cause over-smoothing, obscuring tissue boundaries.

Purpose of the Study:

• To introduce Distance-Weighted Graph Convolutional Networks (DWGCN) for improved spatial domain identification in ST.
• To address limitations of uniform weighting in GCN-based ST analysis.

Main Methods:

• DWGCN replaces uniform neighbor assignment with inverse-distance weighting (IDW).
• Spot-wise normalization and preserved self-loop dominance are employed.
• DWGCN integrates with existing GCN frameworks like SEDR, GraphST, SpaNCMG, and SpaGIC.

Main Results:

• DWGCN integration generally improved clustering accuracy across diverse ST datasets.
• Effectiveness was notable in tissues with complex spatial architectures.
• The method enhanced the preservation of distance-aware spatial structures.

Conclusions:

• DWGCN provides a broadly applicable method for spatial graph analysis in ST.
• The approach improves the delineation of spatial domains by restoring distance-aware structure.