At the beginning of the year when we did the inventory season I have summarized it for you👉 Compilation of common tools for idle deconvolutionThese tools areperformancesHow? What are the choices to be made when doing an analysis? Recently.《Briefings in Bioinformatics" published an overview article reviewing the state-of-the-art methods for ST deconvolution and comprehensively evaluating the performance of 10 methods.
Computational methods developed for deconvolution of cell types for ST data
In recent years, a variety of ST deconvolutionMethods.Existing deconvolution methods for ST data can be broadly categorized into three groups: probabilistic methods, methods based on nonnegative matrix factorization (NMF) and nonnegative least squares (NNLS), and other methods:1) Probabilistic methods, including Adroit, cell2location, DestVI, RCTD, STdeconvolve, and stereoscope , which explicitly or parametrically specify the data distribution and use likelihood-based methods for inference.2) Non-negative matrix factorization andneural networkMethods of decomposition, including spatialDWLS and SPOTlight.3) Other methods includeDSTGand Tangram, using a number of specially designed methodological architectures or lossfunction (math.)to estimate cell type proportions.
* :: All methods except AdRoit are specifically designed for ST data.
Comprehensive test of the performance of ST deconvolution methods
The researchers used ST data from three tissues to assess the performance of the 10 methods described above. For performance quantification they used three metrics: root mean square error (RMSE), distance correlation across cell types, and the difference from true values for each cell type. Smaller RMSE, higher distance correlation, and smaller differences from the true value all indicate better performance.
Mouse olfactory bulb (MOB) data: synthesizing the patterns observed by internal and external references, the results indicate thatRCTD, cell2location and stereoscope are the most robust to batch effects between reference scRNA-seq and target ST data。
Developmental data of the human heart: when deconvolution of pseudo-points constructed from ISS data is performed using an internal reference (i.e., ISS single cells), Adroit, RCTD, stereoscope, DSTG, and Tangram show superior performance, similar to that observed in the MOB data, but here with a much smaller number of genes; when using an external reference, only RCTD and stereoscope were able to capture the expected spatial distribution of cell types. Overall, the methods performed relatively well, except for DestVI and Tangram, which failed to capture the predominant cell types in some layers. Among them.There was high concordance between stereoscope, cell2location and RCTD and ISS cell composition.
Data from primary somatosensory cortical areas in mice:Most methods, especially Tangram and DSTG, achieve excellent performance when using perfectly matched internal references. With internal referencing, Adroit, cell2location, RCTD, and stereoscope still provide satisfactory estimates of cell type proportions despite the limited number of genes available. When using an external reference, RCTD and stereoscope outperform the other methods regardless of the number of genes and platform. Cell2location performs quite well on ST and Visium data when a sufficient number of genes are available.
Attachment: test data sources and references
In summary.RCTD andstereoscopeDemonstrated high stability and accuracy in different organizations.STdeconvolve, as the only reference-free method, has the ability to identify tissue structures and cell mixtures, but cell type mapping must be handled carefully. The scientific team for this review comprehensively evaluated a variety of scenarios, including different tissues, different techniques and data resolutions, different numbers of single cells and spots, and different numbers and types of genes used for analysis. Based on their results, researchers are advised to first identify a number of assessment scenarios that best fit their data, and to select the best performing methods under these scenarios. The selection of reference values, preferably from carefully matched tissue and biological samples, is also essential for deconvolution of ST data. Mismatched scRNA-seq references or references with inaccurately annotated cells can severely impact deconvolution performance. In addition, outside the scope of this review, denoising and downscaling of noisy and high-dimensional ST data can lead to more efficient information extraction. The research team also anticipates that cell-based deconvolution will further benefit from the development and advancement of methods to effectively denoise and reduce the dimensionality of ST data.
We recommend that those interested in the details of the test refer to the original paper ~ or you can visit the following link for details
👉 /JiawenChenn/St-review
bibliography
Chen J, Liu W, Luo T, et al. A comprehensive comparison on cell-type composition inference for spatial transcriptomics data. Brief Bioinform. 2022 Jun 27:bbac245.
Images are from references, please contact to remove any infringement.