.png){kind=icon}
Return
ConMGIN: Interpretable multilayer GIN-Bayesian framework for spatial domain analysis
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neunet.2026.109155.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Accurate identification of spatial domains is crucial for processing spatial transcriptomics in the field of bioinformatics engineering. However, existing methods primarily focus on the similarity of spatial information while neglecting differences in gene expression and tissue structure across distinct spatial locations known as spatial heterogeneity information. To address this issue, this paper proposes ConMGIN, a graph representation learning model constructed using Graph Isomorphism Networks (GIN). ConMGIN integrates GIN with a hybrid Bayesian network and innovatively introduces a Chebyshev-distance-based graph construction that emphasizes the maximum-coordinate deviation to form more robust neighborhood links, thereby enhancing boundary clarity and improving domain separation across heterogeneous tissue structures. This enables more effective modeling of spatial heterogeneity and local relationships. Further interpretability analysis demonstrates that ConMGIN excels in cross-slice clustering and spatial trajectory analysis, particularly in identifying spatial hierarchies and capturing structural similarities across slices. Moreover, an analysis of misclassification cases reveals that ConMGIN maintains high accuracy and consistency when dealing with complex spatial structures. Extensive experiments across multiple spatial transcriptomics datasets spanning different platforms and spatial resolutions show that ConMGIN achieves an ARI improvement of 0.03-0.42 and an NMI improvement of 0.03-0.17 compared to baseline methods, demonstrating strong generalization capabilities and broad applicability.
Keywords:
Spatial transcriptomics
Graph representation learning
Graph contrastive learning
hybrid Bayesian networks
Self-attention mechanism
interpretability
Journal
IF:
6.3
Papers:
8.2K
Citations:
3.0W
Organization
Cited Papers
No cited papers available