SpaGIC: graph-informed clustering in spatial transcriptomics via self-supervised contrastive learning

Abstract

Spatial transcriptomics technologies enable the generation of gene expression profiles while preserving spatial context, providing the potential for in-depth understanding of spatial-specific tissue heterogeneity. Leveraging gene and spatial data effectively is fundamental to accurately identifying spatial domains in spatial transcriptomics analysis. However, many existing methods have not yet fully exploited the local neighborhood details within spatial information. To address this issue, we introduce SpaGIC, a novel graph-based deep learning framework integrating graph convolutional networks and self-supervised contrastive learning techniques. SpaGIC learns meaningful latent embeddings of spots by maximizing both edge-wise and local neighborhood-wise mutual information of graph structures, as well as minimizing the embedding distance between spatially adjacent spots. We evaluated SpaGIC on seven spatial transcriptomics datasets across various technology platforms. The experimental results demonstrated that SpaGIC consistently outperformed existing state-of-the-art methods in several tasks, such as spatial domain identification, data denoising, visualization, and trajectory inference. Additionally, SpaGIC is capable of performing joint analyses of multiple slices, further underscoring its versatility and effectiveness in spatial transcriptomics research.

Keywords: graph convolutional networks; self-supervised contrastive learning; spatial domain identification; spatial transcriptomics.

Figure 1

The overall framework of SpaGIC. (A) SpaGIC first constructs a feature graph , where the node connection relationships are calculated by KNN based on spatial location similarity, and node features are derived from preprocessed gene expression data. Next, SpaGIC employs a GCN-based auto-encoder to learn the node latent embeddings . (B) These latent representations are then used to reconstruct the connection weights between nodes and establish mutual information constraints on the graph structures. (C) Calculating an InfoNCE-like loss based on the learned embeddings introduces a contrastive learning constraint. (D) A decoder reverses the embeddings back into the original feature space to reconstruct the gene expression matrix . (E) The output of SpaGIC can be applied to various ST downstream analysis tasks, such as spatial clustering, data denoising, visualization, trajectory inference, and multi-slice joint analysis.

Figure 2

Spatial domains identification and data denoising on the DLPFC dataset. (A) Manual annotation of the DLPFC 151673 slice. (B) ARI boxplots of eight methods on 12 DLPFC slices. In the boxplot, the center line denotes the median, box limits denote the upper and lower quartiles, and whiskers denote the 1.5 interquartile range. (C) The spatial domains identified by Scanpy, SpaGCN, DeepST, SEDR, STAGATE, Spatial-MGCN, GraphST, and SpaGIC on the DLPFC 151673 slice. (D) UMAP visualization and PAGA graph generated based on the embedding by these methods on the 151673 slice. (E) Visualization of the raw expression of layer marker genes in the 151673 slice, both before and after denoising by SpaGIC.

Figure 3

Spatial domains identification on the Stereo-seq MOB dataset. (A) The laminar structure of the Stereo-seq MOB annotated in DAPI-stained images. (B) Clustering results from Scanpy, SpaGCN, DeepST, STAGATE, GraphST, and SpaGIC. (C) Visualization of spatial domains detected by SpaGIC and related marker gene expression.

Figure 4

Spatial domains identification on the STARmap MVC dataset. (A) Manual annotation of the STARmap MVC. (B) ARI and NMI bar charts for eight methods. (C) The spatial domains identified by Scanpy, SpaGCN, DeepST, SEDR, STAGATE, Spatial-MGCN, GraphST, and SpaGIC. (D) UMAP visualization and PAGA graph generated by the embedding of these methods.

Figure 5

Joint analysis on the DLPFC dataset. (A) Aligned spatial domain identified by Harmony, STAGATE, SEDR, and SpaGIC via joint analysis of four slices of sample 3 (151673-151676). (B) UMAP visualization of embeddings colored by slices (top), ground truth (middle), and identified domains (bottom).

Figure 6

The ARI boxplots of SpaGIC and its variants on the DLPFC dataset.