Edit |Violet.
Spatial transcriptomics (ST) techniques detect mRNA expression in individual cells while preserving their two-dimensional (2D) spatial coordinates, enabling researchers to study the spatial distribution of the transcriptome in tissues;However, it remains a challenge to perform joint analysis of multiple ST slices and align them to build a three-dimensional (3D) stack of tissues.
Recently, a research team from the University of Science and Technology of China, Hefei National Science Center, and Beijing Institute of Biological Sciences (NIBS) proposed a deep Xi spatial architecture representation (SPACEL) for ST data analysis. SpaceL consists of three modules – Spoint, Splane, and Scube – that cover three analytical tasks for ST data.
Comparison with 19 state-of-the-art methods using simulated and real ST datasets and ST techniques from a variety of organizations demonstrates that SPACEL outperforms other methods in terms of cell type deconvolution, spatial domain recognition, and 3D alignment, thus demonstrating SPACEL as a valuable integrated toolkit for ST data processing and analysis.
The study, titled "SpaceL: Deep Learning-Based Characterization of Spatial Transcriptome Architectures," was published in Nature Communications on November 22, 2023.
Spatial transcriptomics (ST) techniques enable researchers to examine the spatial distribution of the entire transcriptome in histological sections in principle, greatly improving our understanding of organ structure and disease microenvironment. There are two broad classes of experimental ST techniques that can i) detect the expression of a portion of the transcriptome at single-cell resolution, or ii) examine the entire transcriptome without single-cell resolution.
The current trend of rapid development and application of ST technology suggests that a large number of ST slices from different tissues (and conditions) will be generated in scientific and medical research for the foreseeable future. Therefore, there is an urgent need for computational tools that can quickly and efficiently implement the integrated analysis of multiple ST slice data.
Accurately identifying functional spatial domains across multiple slices and reconstructing the 3D structure of tissues provides valuable opportunities for significant biological discoveries in a variety of real-world applications. Overall, stacked 3D alignment of structuring organizations remains a significant challenge for ST datasets with significant flaws.
Here, the researchers developed SpaceL, a toolkit based on deep Xi. SpaceL consists of three modules: Spoint embeds a multilayer perceptron with a probabilistic model for deconvolution of the cell type composition at each point in a single ST slice;Splane employs a graph convolutional network approach and an adversarial Xi algorithm to identify spatial domains that are transcriptome- and spatially consistent across multiple ST slices;The Scube automatically converts the spatial coordinate systems of serial slices and stacks them together to build the 3D structure of the organization.
Figure 1: The spacel process. (*To ensure the robustness of SpaceL, the researchers conducted extensive experiments to evaluate its performance at various hyperparameter settings.) The results show that Spoint, Splane, and Scube exhibit superior robustness to hyperparameter changes compared to other state-of-the-art methods, emphasizing the effectiveness of SpaceL in providing reliable and consistent results across different experimental setups.
The researchers used SpaceL to analyze 11 ST datasets, including 156 slices acquired using 10x Visium, StarMap, Merfish, Stereo-Seq, and Spatial Transcriptomics technologies. SpaceL outperforms other state-of-the-art methods in all three examined analytical tasks and therefore represents a valuable integrated toolkit for ST data processing and analysis.
Of the 11 deconvolution methods, Spoint yielded the highest average PCC SSIM value (= 0.).73/0.69), and the lowest average RMSE JSD value (=0.).05/0.41）。In addition, the accuracy score (AS) defined in the benchmark study was applied to evaluate the performance of each method: the mean AS ( = 093) significantly higher than other methods (as = 0.).24-0.82）。
Figure 2: Cell-type deconvolution of points using SPOINT and other deconvolution methods. (*After deconvolution of the cell type using Spoint, Splane is applied to identify the spatial domain of the ST slice of the DLPFC dataset described above.) The results show that the joint analysis scheme improves the accuracy of Splane spatial recognition.
Figure 3: Spatial domains were identified from 12 10x visium slices of DLPFC. (*Immediately after that, the performance of Splane in identifying the spatial domain of disease slices was tested.) The results demonstrate Splane's ability to identify consistent tumor regions and boundaries across multiple sections of the tumor system.
Figure 4: Joint analysis of 11 breast cancer ST sections from three different datasets. (*Researchers have also developed SpaceL's Scube module to build and study the 3D architecture of a given organization.) SpaceL's Scube module outperforms Staligner and Paste in both simulated (Starmap) and real-world (Merfish and Stereo-Seq) dataset alignment and 3D architecture construction.
Figure 5: 3D alignment of serial ST slices of mouse brain. (*To highlight the integrated nature of SpaceL, the researchers applied a fully integrated workflow to analyze mouse whole-brain ST data.) Studies have shown that SpaceL can be an effective integrated toolkit for analyzing multiple ST slices.
*Links: