Path-MGCN: a pathway activity-based multi-view graph convolutional network for determining spatial domains

Abstract

Spatial transcriptomics (ST) comprehensively measure the gene expression profiles while preserving the spatial information. Accumulated computational frameworks have been proposed to identify spatial domains, one of the fundamental tasks of ST data analysis, to understand the tissue architecture. However, current methods often overlook pathway-level functional context and struggle with data sparsity. Therefore, we develop Path-MGCN, a multi-view graph convolutional network (GCN) with attention mechanism, which integrates pathway information. We first calculate spot-level pathway activity scores via gene set variation analysis from gene expression and construct distinct adjacency graphs representing spatial and functional proximity. A multi-view GCN learns spatial, pathway, and shared embeddings adaptively fused by attention and followed by a Zero-inflated negative binomial decoder to retain the original transcriptome information. Comprehensive evaluations across diverse datasets (human dorsolateral prefrontal cortex, breast cancer and mouse brain) at various resolution demonstrate Path-MGCN's superior accuracy and robustness, significantly outperforming state-of-the-art methods and maintaining high performance across different pathway databases (Kyoto Encyclopedia of Genes and Genomes, Gene Ontology, Reactome). Crucially, Path-MGCN enhances biological interpretability, enabling the identification of Tertiary lymphoid structure-like regions and spatially resolved metabolic heterogeneity (hypoxia, glycolysis, AMP-activated protein kinase signaling) linked to tumor progression stages in human breast cancer. By effectively integrating functional context, Path-MGCN advances ST analysis, providing an accurate and interpretable framework to dissect tissue heterogeneity and enables detailed spatial mapping of molecular pathways that highlights potential targeted therapeutic strategies crucial for developing safe and effective synergistic anti-tumor therapies.

Keywords: attention mechanism; multi-view graph convolutional network; pathway; spatial domain; spatial transcriptomics.

Figure 1

Overview of Path-MGCN. Path-MGCN takes input of gene expression matrix and spatial coordinates from ST datasets. Path-MGCN calculates a pathway activity matrix by GSVA based on pathway databases (KEGG, GO, and Reactome) using the expression of highly variable genes. The adjacency graphs representing spatial proximity and pathway activity similarity are constructed. Path-MGCN then applies a multi-view GCN to learn distinct latent representations for spatial, pathway and shared features. These embeddings are adaptively fused via an attention mechanism into a unified representation, which is subsequently decoded by a ZINB framework. The final embedding is utilized for spatial clustering to identify domains, followed by visualization and downstream biological interpretation, such as identifying domain-specific pathways.

Figure 2

Path-MGCN achieves outstanding performance in identifying spatial domains. (A) Manual annotation for the DLPFC slice 151,507. (B) Boxplot of ARI scores for eight methods on 12 DLPFC slices. (C) the clustering results and the UMAP visualization generated by path-MGCN, MManiST, spatial-MGCN, SEDR, STAGATE, SpaGCN, stLearn and stDCL on slice 151,507. (D) Lollipop plots displaying the activity of seven important domain-specific pathways in each spatial domain identified by path-MGCN. (E) Dot plot of the pathway activity of domain-specific pathways. Red and blue colours indicate relatively higher (positive) and lower (negative) normalized activity, respectively. The circle size is proportional to the magnitude of that normalized activity. (F) Activity visualization of specific pathways for domain 0, domain 1 and domain 6.

Figure 3

Path-MGCN achieves higher accuracy in identifying cancer tissue heterogeneity of human breast cancer. (A) Manual annotation on 10x Visium data of human breast cancer sample BAS1. (B) The clustering result on BAS1 generated by Path-MGCN. Numbered domains in the figure correspond to those in the manual annotation. (C) The clustering results on BAS1 generated by Spatial-MGCN, stLearn, STAGATE, and SEDR. (D) Manual annotation on human breast cancer sample DCIS. (E) The clustering results on DCIS generated by Path-MGCN, Spatial-MGCN and stLearn. (F) Manual annotation on human breast cancer sample IDC. (G) The clustering results on IDC generated by Path-MGCN, Spatial-MGCN and stLearn.

Figure 4

Path-MGCN uncovers pathways’ spatial patterns to further dissect cancer heterogeneity of human breast cancer. (A) Heatmap of Pearson correlation coefficient between domains. The correlation coefficient is calculated based on the median pathway activity of all spots within a domain. The morphotypes annotation on the left side of the heatmap are obtained by mapping our clustering results onto the manual annotations. (B) Dot plot of the pathway activity of domain-specific pathways. The color intensity represents the median pathway activity of domain-specific pathways, and the dot size represents the median normalized pathway expression of domain-specific pathways. Activity visualization of specific pathways for domain 18 (C), domain 0 (D), and domain 5 (E). The red line areas correspond to the manual annotation areas of DCIS/LCIS_1, IDC_3 and IDC_4, respectively. Box plot of median activity (F) and median normalized expression (G) of domain-specific pathways between DCIS/LCIS and IDC morphotypes.

Figure 5

Path-MGCN achieves better clarity in laminar structure identification. (A) Laminar organization of the Stereo-seq MOB annotated by DAPI-stained images. (B) The clustering result of STAGATE, SEDR, Spatial-MGCN, and Path-MGCN. (C) Visualization of spatial domains generated by Path-MGCN (upper) and the corresponding marker gene expressions (lower). (D) Laminar organization of the slide-seq V2 MOB annotated by the Allen Reference Atlas. (E) The clustering results on slide-seq V2 MOB generated by Path-MGCN, Spatial-MGCN, STAGATE, and SEDR. (F) Manual annotation on STARmap data of mouse visual cortex. (G) The clustering result on mouse visual cortex of Path-MGCN. (H) Manual annotation on osmFISH data of mouse somatosensory cortex. (I) The clustering result on mouse somatosensory cortex of Path-MGCN.

Figure 6

Ablation study and parameter analysis of Path-MGCN. (A) ARI scores for seven distinct ablation experiments on 12 DLPFC slices. (B) ARI scores for Path-MGCN versus the number of KNN neighbors (k) on four datasets.