【Nature review Genetis】最新报道，就算一天一种反卷积方法，两个月都写不完！

空间转录组学是一种研究细胞空间组织结构的强大方法，而细胞的空间组织是多细胞生命在发育、功能和演化中的关键特征。然而，基于测序的空间转录组学尚未达到单细胞分辨率，因此需要先进的解卷积（deconvolution）方法来推断每个空间位置上的细胞类型贡献。近年来的进展催生了多种细胞类型解卷积工具，帮助研究者描述健康和疾病状态下的组织结构。然而，层出不群的工具让研究者目不暇接，根本不知道有什么，选择什么，学习方法的速度赶不方法的出现。这里介绍一篇Nature子刊，系统的介绍了目前的反卷积方法，好家伙，数了一下，足足有70多种，每天学一个，两个月干不完！

文章链接：Gaspard-Boulinc, L.C., Gortana, L., Walter, T. et al.Cell-type deconvolution methods for spatial transcriptomics. Nat Rev Genet26, 828–846 (2025). https://doi.org/10.1038/s41576-025-00845-y

这篇文献总结的很详细，值得阅读，这个帖子不是对文献的翻译，抛出两个问题，也是很多人最关心的。第一个是方法有哪些？第二个是哪个方法好？

一、到底有哪些方法！

按照模型、是否需要参考数据集，进行了详细的分类！可以对应自己用过哪些，比如我们已经介绍的RCTD，SPOTlight都是需要reference，但都是基于贝叶斯模型框架。

figure 2

Methods are grouped by the frameworks used to perform deconvolution. Bayesian models are divided into three subtypes (blue). The first is topic modelling and the other two correspond to two distribution assumptions used in classical Bayesian frameworks — negative binomial and Poisson (or Poisson-gamma) distributions. GIST has its own branch, as it uses a t-distribution. Regression models also fall into several categories, differing from one another by, for example, data assumptions and imposed constraints (yellow). Optimal-transport-based methods are grouped into a single category (green), as are dimensionality reduction methods, which are mainly based on matrix factorization (purple), except for LANTSA (landmark-based). EnDecon stands alone in its category as the only consensus method among the other deconvolution methods (grey). Finally, most methods belong to the deep-learning-based category, with a subcategory dedicated to graph-based networks (red). The other deep learning architectures use autoencoders, fully connected neural networks or others to perform deconvolution. *Reference-free methods. **Both reference-free and reference-based implementations; methods that can be used with or without a reference. DWLS, dampened weighted least squares; NNLS, non-negative least squares.

二、用哪个方法好？

哪种方法最好，这也是很多人纠结用什么的问题，我想没有定论，同一种方法，不同的数据进行测试，可能得到不一样的效果。故而，在实际的分析中，不妨测试1～2种，已经在文章中大量测试使用的方式，可以有效节约自己的时间，并得到较好的结果。

figure 3

Summary of deconvolution method characteristics across six categories: use of reference, use of extra modality (coordinates and/or image), programming languages according to GitHub repositories, output types, class of framework following our proposed classification and presence in independent benchmarks . Methods are grouped according to shared characteristics. Method names in grey are methods not yet peer reviewed. Left: decision tree to identify all possible methods in function of the data types available and the modalities the user wants to use.

最后，作者也说出了结论：用于细胞类型解卷积推断的方法在优势、假设和局限性上各不相同，在应用这些策略以及选择最合适的分析参数时，理解这些差异至关重要。