来源于 Benchmarking single-cell RNA-sequencing protocols for cell atlas projects
对比了13 commonly used scRNA-seq and single-nucleus RNA-seq的方法对比，也算是各有千秋。结果来看，不尽相同。在采用方法策略时候，还是要结合自己的课题，选择合适的方法，不能乱来。
摘要
单细胞RNA测序(scRNA-seq)是一项用于分辨样本中单细胞水平转录组的领先技术。最新的一些protocols可hold住成千上万级别单细胞的测序，并已经被用于展示组织器官和生物体水平的cell atlases。然而，这些不同的protocols在RNA捕获效率、捕获偏倚程度、单细胞规模和建库成本方面存在很大差异，它们在不同应用方向中的相对优劣性尚不很清楚。
本研究生成了一个基准数据集，用以系统地评估这些单细胞测序的protocols在全面单细胞类型分辨能力和状态方面的能力。我们进行了一项多中心研究，用混合多种细胞的异质参考样本，对13种常用的scRNA-seq和单核RNA-seq protocol 进行了评测。比较分析显示各个protocols性能有显著差异。这些protocols在文库的复杂性和检测细胞类型markers的能力上有所不同，这些指标影响了它们的预测值和整合到 reference cell atlases的普适性。本结果为研究人员和联合项目(如人类细胞图谱Human Cell Atlas)提供了指导守则。
Abstract
Single-cell RNA sequencing (scRNA-seq) is the leading technique for characterizing the transcriptomes of individual cells in a sample. The latest protocols are scalable to thousands of cells and are being used to compile cell atlases of tissues, organs and organisms. However, the protocols differ substantially with respect to their RNA capture efficiency, bias, scale and costs, and their relative advantages for different applications are unclear.
In the present study, we generated benchmark datasets to systematically evaluate protocols in terms of their power to comprehensively describe cell types and states. We performed a multicenter study comparing 13 commonly used scRNA-seq and single-nucleus RNA-seq protocols applied to a heterogeneous reference sample resource. Comparative analysis revealed marked differences in protocol performance. The protocols differed in library complexity and their ability to detect cell-type markers, impacting their predictive value and suitability for integration into reference cell atlases. These results provide guidance both for individual researchers and for consortium projects such as the Human Cell Atlas.
测试样品是一个包含人、鼠、狗细胞的混合细胞样品，用于测试13种单细胞测序方案。获得的reads分别mapping到人、鼠、犬的参考序列上，分别计算不同物种的不同测序方法的基因表达量。
The reference sample consists of human PBMCs (60%), and HEK293T (6%), mouse colon (30%), NIH3T3 (3%) and dog MDCK cells (1%). The sample was prepared in one single batch, cryopreserved and sequenced by 13 different sc/snRNA-seq methods. Sequences were uniformly mapped to a joint human, mouse and canine reference, and then separately to produce gene expression counts for each sequencing method.
a, Color legend of sc/snRNA-seq protocols.
b, 人细胞UMAP of 30,807 cells from the human reference sample (Chromium) colored by cell-type annotation.
c, 鼠细胞UMAP of 19,749 cells from the mouse reference (Chromium) colored by cell-type annotation.
d, Boxplots displaying the minimum, the first, second and third quantiles, and the maximum number of genes detected across the protocols, in down-sampled (20,000) HEK293T cells, monocytes and B cells. Cell identities were defined by combining the clustering of each dataset and cell projection on to the reference.
e, Number of detected genes at stepwise. down-sampled, sequencing depths. Points represent the average number of detected genes as a fraction of all cells of the corresponding cell type at the corresponding sequencing depth.
f, Dropout probabilities as a function of expression magnitude, for each protocol and cell type, calculated on down-sampled data (20,000) for 50 randomly selected cells.
a,b, Principal component analysis on down-sampled data (20,000) using highly variable genes between protocols, separated into HEK293T cells, monocytes and B cells, and color coded by protocol (a) and number of detected genes per cell (b).
c, Pearson’s correlation plots across protocols using expression of common genes. For a fair comparison, cells were down-sampled to the same number for each method (B cells, n?=?32; monocytes, n?=?57; HEK293T cells, n?=?55). Protocols are ordered by agglomerative hierarchical clustering.
d, Average log(expression) values of cell-type-specific reference markers for down-sampled (20,000) HEK293T cells, monocytes and B cells.
e, Log(expression) values of reference markers on down-sampled data (20,000) for HEK293T cells, monocytes and B cells (maximum of 50 random cells per technique).
f, Cumulative gene counts per protocol as the average of 100 randomly sampled HEK293T cells, monocytes and B cells, separately on down-sampled data (20,000).
a, The tSNE visualizations of unsupervised clustering in human samples from 13 different methods. Each dataset was analyzed separately after down-sampling to 20,000?reads?per cell. Cells are colored by cell type inferred by matchSCore2 before down-sampling. Cells that did not achieve a probability score of 0.5 for any cell type were considered unclassified.
b, Clustering accuracy and ASW for clusters in each protocol.
a–d, UMAP visualization of cells after integrating technologies for 18,034 human (a,b) and 7,902 mouse (c,d) cells. Cells are colored by cell type (a,c) and sc/snRNA-seq protocol (b,d).
e,f, Barplots showing normalized and method-corrected (integrated) expression scores of cell-type-specific signatures for human HEK293T cells, monocytes, B cells (e), and mouse secretory and TA cells (f). Bars represent cells and colors methods.
g,h, Evaluation of method integratability in human (g) and mouse (h) cells. Protocols are compared according to their ability to group cell types into clusters (after integration) and mix with other technologies within the same clusters. Points are colored by sequencing method.
Methods are scored by key analytical metrics, characterizing protocols according to their ability to recapitulate the original structure of complex tissues, and their suitability for cell atlas projects. The methods are ordered by their overall benchmarking score, which is computed by averaging the scores across metrics assessed from the human datasets.
参考文献：
- Analysis
- Published: 2020-04-06
Benchmarking single-cell RNA-sequencing protocols for cell atlas projects
Elisabetta Mereu, Atefeh Lafzi Holger Heyn*
Nature Biotechnology volume 38, pages747–755(2020)Cite this article
