用米氏方程解决单细胞转录组dropout现象
米氏方程(Michaelis-Menten equation): v=Vmax × [S] /(Km+[S])
在假定存在一个稳态反应条件下推导出来的,其中 Km 值称为米氏常数,Vmax是酶被底物饱和时的反应速度,[S]为底物浓度。
Km值的物理意义为反应速度(v)达到1/2Vmax时的底物浓度(即Km=[S]),单位一般为mol/L,只由酶的性质决定,而与酶的浓度无关。可用Km的值鉴别不同的酶。
今天要介绍的这篇文章提出了一个算法,R包是:M3Drop , 文章是:Modelling dropouts for feature selection in scRNASeq experiments
(值得一提的是,该文章发表于预印本,不是正式SCI期刊)
挑选重要基因
目前已有的寻找单细胞转录组测序数据中的重要基因(feature selection)的方法都不够好,比如 scLVM 主要是根据先验基因集,比如cell-cycle or apoptosis来区分细胞。与此相反,基于 highly variable genes (HVG) 的方法挑选到的变化量大的那些基因很可能是技术带来的误差。而且低表达量基因的变动往往大于高表达量基因,而且所谓的表达变化大也并没有很好的生物学解释。
一个比较好理解的概念是差异基因,但是需要预先把细胞群体分组后进行比较才能得到,而很多时候细胞太相似了,没办法很好的分开。像PCA或者t-SNE这样的降维方法也可以用来挑选重要基因,但它们也受制于系统误差或者批次误差等等。
dropout是scRNASeq数据的一大特点,就是很多基因在某些细胞根本就不表达,但是在另外的细胞却高表达。这篇文章作者对全长转录本数据和基于UMI的表达量数据分别提出了对应的解决方案,Michaelis-Menten equation 和 depth adjusted negative binomial (DANB)
单细胞转录组数据里面的dropouts可以达到50%,但是通常认为这个dropouts是因为在文库构建的过程中,有部分基因没有被成功的反转录,是一个酶促反应。
所以作者用Michaelis-Menten 来建模。
比较了9种 feature selection 方法,
使用它们分别对基因排序,算法如下:
by the magnitude of their loadings in principal component analysis (PCA)
by the strength of their most negative gene-gene correlation (Cor)
by their relative Gini index (Gini)
M3Drop dropouts-mean expression curve (M3Drop)
the squared coefficient of variation (CV2)
mean expression relationship (HVG)
the dispersion-mean expression relationship fit by DANB (NBDisp)
the dropouts-mean expression relationship fit by DANB (NBDrop).
这些算法都不需要预先对样本进行分类,是无监督的算法。
differentially variable (DV) genes
highly variable (HV) genes
differentially expressed (DE) genes
单细胞转录组数据的batch effects比较严重,所以 feature selection 过程的一个主要目的就是降低技术误差的影响,集中在有生物学意义的差异上面。
公共数据集
作者比较了 5个公共数据集,都是小鼠的胚胎细胞,含有17~255个细胞的测序数据,包括zygote to blastocyst.
Tung et al. (2017) [12] considered iPSCs from three different individuals and performed three replicates of UMI-tagged scRNASeq and three replicates of bulk RNASeq for each. (GSE77288 ).
For Kolodziejczyk et al. (2015),we considered ESCs grown under two conditions: alternative 2i and serum for which there were three replicates of scRNASeq and two replicates of bulk RNASeq.( E-MTAB-2600 )
对bulk转录组数据用了3种方法找差异基因,分别是 DESeq2,edgeR,limma-voom只有3种方法都是 5% FDR的差异基因才认为是阳性标准基因集,那些3种方法都在 20% FDR的非差异基因认为是阴性金标准。
1,915 positives, and 8,398 negatives for the iPSCs
709 positives and 11,278 negatives for the ESCs
有了这些基因,就可以计算ROC
单细胞转录组数据文章一般分成下面两大类:
第一大类是:deep sequencing of full-transcripts for a relatively small number of cells
代表性的文章如下:
Accounting for technical noise in singlecell RNAseq experiments. Nat. Methods 10, 1093–1095 (2013).
Fast, scalable and accurate differential expression analysis for single cells. (2016). doi:10.1101/049734
Singlecell RNAseq reveals dynamic, random monoallelic gene expression in mammalian cells. Science 343, 193–196 (2014). 14. Brennecke, P. et al. Accounting for technical noise in singlecell RNAseq experiments. Nat. Methods 10, 1093–1095 (2013).
Dynamics of Global Gene Expression Changes during Mouse Preimplantation Development. Dev. Cell 6, 117–131 (2004).
Roles of CDX2 and EOMES in human induced trophoblast progenitor cells. Biochem. Biophys. Res. Commun. 431, 197–202 (2013).
第二类是:high-cell number, low-depth sequencing of 3’ or 5’ ends of transcripts tagged with unique molecular identifiers
代表性的文章是:
Quantification noise in single cell experiments. Nucleic Acids Res. 39, e124 (2011).
Quantification of mRNA in single cells and modelling of RTqPCR induced noise. BMC Mol. Biol. 9, 63 (2008).
ZIFA: Dimensionality reduction for zeroinflated singlecell gene expression analysis. Genome Biol. 16, 241 (2015).
DNA methylation dynamics during epigenetic reprogramming in the germline and preimplantation embryos. Genes Dev. 28, 812–828 (2014).
Genetic programs in human and mouse early embryos revealed by singlecell RNA sequencing. Nature 500, 593–597 (2013).