Spatial transcriptome technology generates gene expression profiles with spatial context and requires spatial information analysis tools to accomplish three key tasks: spatial聚类、多样本整合和细胞类型去卷积。近日，《Nature Communications》发表了一种图自我监督的对比学习方法：GraphST，其充分利用空间转录组学数据，以优于现有方法。
GraphST是什么？
GraphST是一种图自我监督对比学习方法，它充分利用空间信息和基因表达谱进行空间信息聚类、整合和细胞类型去卷积。通过在GraphST中使用自我监督对比学习，发现它提高了学习下游分析的相关潜在特征的性能。
GraphST包括三个模块，Each module features a graphical self-supervised comparative learning architecture customized for each of the three tasks：Spatially informative clustering (top panel A), vertical and horizontal batch integration of multiple tissue slices (top panel B), and spatial cell type deconvolution by projection of scRNA-seq to ST (top panel C).. In all three modules, using spatial transcriptomicsdata setof spatial information to construct a neighborhood graph in which spatially close points to each other are connected. Next, a graph convolutional network is constructed as an encoder to embed gene expression profiles and spatial similarities into the latent representation space by iteratively aggregating gene expressions from neighboring points.
Performance testing of GraphST
The development team conducted extensive testing using GraphST on three analysis tasks on different 10x Visium, Stereo-seq and Slide-seqV2 datasets of human and mouse tissues, including human brain, human breast cancer tissue, human lymph nodes, mouse breast cancer, mouse olfactory bulb, mouse brain and mouse embryo.
Clustering tests show that GraphST outperforms seven existing methods in recognizing spatial domains.
GraphST clustering improves the identification of tissue structures in human dorsolateral prefrontal cortex (DLPFC), mouse olfactory bulb and mouse hippocampal tissue.
GraphST is able to accurately identify different organs in Stereo-seq mouse embryos.
Joint analysis of mouse breast cancer and mouse brain datasets demonstrates that GraphST is able to accurately identify spatial domains from multiple tissue sections while effectively eliminating batch effects without explicitly detecting batch factors.
GraphST is capable of accurate vertical and horizontal integration of mouse breast cancer ST data and mouse anterior and posterior brain data, respectively.
The development team also tested GraphST's projection of scRNA-seq data onto ST to predict cell status (cell type and sample type) in spatial points. The computed cell point mapping matrices estimated cell type composition more accurately than cell2location (the best performing deconvolution method).
Comparison of the accuracy of GraphST with the TOP deconvolution method cell2location in predicting the spatial distribution of scRNA-seq data versus simulated data, human lymph nodes, and DLPFC slices 151673.
GraphST can transfer the phenotype of samples extracted in scRNA-seq to ST. The development team demonstrated this ability by depicting tumor and normal adjacent regions in tumor-derived tissue sections.
GraphST enables comprehensive and accurate spatial mapping of scRNA-seq data from human breast cancer data.
GraphST Toolkit'sexpand one's financial resourcesThe Python implementation is available at the following link:
/JinmiaoChenLab/GraphST.
bibliography
Long, Y., Ang, ., Li, M. et al. Spatially informed clustering, integration, and deconvolution of spatial transcriptomics with GraphST. Nat Commun 14, 1155 (2023).