DeepSpaCE: A deep learning framework to detect spatial single cell domains
The rapidly growing field of spatial transcriptomics has provided new means to discover cellular states within their local microenvironments, thereby revealing the spatial structures that form the basis of organ function, development, and disease pathology (Sudupe, Laura, et al., 2023). Cells that are spatially close to each other are often deemed to be functionally related, emphasizing the need to consider spatial relationships to understand cellular roles in the tissue. However, despite significant advances in spatial molecular imaging technologies, such as MERFISH and 10X Xenium, current methodologies face considerable challenges in accurately deciphering spatial domains and cell niches. These challenges stem from the need for analytical frameworks capable of integrating these data dimensions to extract meaningful biological insights.
Traditional approaches in spatial omics have often adapted clustering algorithms initially designed for single-cell RNA sequencing, that take into account only the cell transcriptomic profiles and overlook spatial information. Moreover, recent developments have introduced spatially informed algorithms, based on deep learning methods (Liu, Teng, et al., 2024; Long, Yahui, et al., 2023) and Bayesian methods (Li, Zheng, and Xiang Zhou., 2022). However, these tools encounter challenges in robustness and generalizability across various tissue types. Moreover, they usually have multiple hyperparameters that their selection remains, to a large degree, user-based and non-trivial.
In this work, we introduce DeepSpaCE, a Deep learning-based Spatial Cell Explorer model, that employs graph neural networks to address the challenge of identifying spatial domains and cell niches with high accuracy and robustness. We summarize the significance and novelties of our work in the following.
-
Leveraging a GNN-based encoder-decoder architecture and bi-objective function, DeepSpaCE takes into account both cell transcriptomic data and spatial relationships among cells.
-
DeepSpaCE requires only three hyperparameters. We either demonstrate the robustness of our pipeline to the choice of a hyperparameter or provide algorithmic solutions to find the optimum hyperparameter.
We applied DeepSpaCE on spatial single-cell data and showed the robustness of our method in accurately detecting the spatial domains and its superiority over existing methods. Additionally, we utilized DeepSpaCE to identify spatial domains in in-house kidney biopsies obtained using 10X Xenium technology. The results demonstrated the potential of DeepSpaCE to significantly enhance the accuracy of spatial domain detection in diverse tissue types, thereby providing a means to more precise and insightful spatial transcriptomic analyses.
References
Li, Zheng, and Xiang Zhou. "BASS: multi-scale and multi-sample analysis enables accurate cell type clustering and spatial domain detection in spatial transcriptomic studies." Genome biology 23.1 (2022): 168.
Liu, Teng, et al. "A comprehensive overview of graph neural network-based approaches to clustering for spatial transcriptomics T. Liu et al. Overview of Spatial Transcriptomics’ Spatial Clutering." Computational and Structural Biotechnology Journal (2023).
Long, Yahui, et al. "Spatially informed clustering, integration, and deconvolution of spatial transcriptomics with GraphST." Nature Communications 14.1 (2023): 1155.
Sudupe, Laura, et al. "Spatial Transcriptomics Unveils Novel Potential Mechanisms of Disease in a MI cγ1 Multiple Myeloma in vivo Model." Blood 142 (2023): 4076.
