stMMR: accurate and robust spatial domain identification from spatially resolved transcriptomics with multimodal feature representation

Abstract

Background: Deciphering spatial domains using spatially resolved transcriptomics (SRT) is of great value for characterizing and understanding tissue architecture. However, the inherent heterogeneity and varying spatial resolutions present challenges in the joint analysis of multimodal SRT data.

Results: We introduce a multimodal geometric deep learning method, named stMMR, to effectively integrate gene expression, spatial location, and histological information for accurate identifying spatial domains from SRT data. stMMR uses graph convolutional networks and a self-attention module for deep embedding of features within unimodality and incorporates similarity contrastive learning for integrating features across modalities.

Conclusions: Comprehensive benchmark analysis on various types of spatial data shows superior performance of stMMR in multiple analyses, including spatial domain identification, pseudo-spatiotemporal analysis, and domain-specific gene discovery. In chicken heart development, stMMR reconstructed the spatiotemporal lineage structures, indicating an accurate developmental sequence. In breast cancer and lung cancer, stMMR clearly delineated the tumor microenvironment and identified marker genes associated with diagnosis and prognosis. Overall, stMMR is capable of effectively utilizing the multimodal information of various SRT data to explore and characterize tissue architectures of homeostasis, development, and tumor.

Keywords: domain identification; geometric deep learning; multimodal integration; similarity contrastive learning; spatially resolved transcriptomics.

Figure 1:

Schematic overview of stMMR for the joint representation of features from different modalities. Gene expression and histology image information are embedded using the GCN module based on the adjacent matrix. Then, the relationships between different modalities are captured through similarity contrast learning, followed by feature fusion. Finally, the original features are reconstructed from the multimodal feature representation. This representation can be used for downstream analysis directly.

Figure 2:

Performance comparisons of different methods on DLPFC datasets. (A) The histology image and manually annotated brain regions of slice 151509. (B) The overall performance of 8 different methods across 12 slices. (C) The domain recognition results on slice 151509. (D) The UMAP visualization results of the embeddings from 8 different methods on slice 151509. (E) The inferenced trajectories on slice 151509.

Figure 3:

stMMR enhances spatial gene expression profiles and spatial structural characterization. (A) Spatial representation of layer-specific marker genes before and after data enhancement. (B) Gene expression level before and after data enhancement.

Figure 4:

stMMR reveals cell lineage structures during chicken heart development. (A) The ground-truth label provided by the original data. (B) The domains recognized by stMMR and SpaceFlow. (C) The plots of pSM value from stMMR and SpaceFlow for illustrating the pseudo-temporal developmental trajectory. (D) The differentially expressed marker genes discovered by stMMR.

Figure 5:

stMMR identifies tumor region in the human breast cancer dataset. (A) The H&E images and the manually annotated regions. (B) The annotation results from different methods. (C) Top 20 differentially expressed gene enriched terms identified by DisGeNET (left panel) and TRRUST (right panel).

Figure 6:

stMMR recognizes cell-type differences in a lung cancer dataset. (A) One FOV of the lung cancer SRT data. (B) Cell types identified by different methods. (C) Expression pattern of marker genes for different cell types. (D) The overall performance of different methods across 20 FOVs. (E) Cell types annotated manually in 20 FOVs. And cell types annotated by stMMR in 20 FOVs. (F) The zoomed-in results of boundaries between adjacent FOVs identified by stMMR.