6、Marker gene detection

原文链接Chapter 11 Marker gene detection

1、目的

• we identify the genes that drive separation between clusters. 鉴别直接导致分群结果的marker gene.
• In the most obvious case, the marker genes for each cluster are a priori associated with particular cell types .即 marker gene 与后面的细胞类型注释直接相关
• different approaches can be used to consolidate test results into a single ranking of genes for each cluster.
• 本文主要学习scran包的findmarkers函数其中的t-test方法。

load("fluidigm.clust.RData") 
#加载上一步分析结果
fluidigm.clust
library(scran)  

2、findMarkers

（1）基础用法

markers.fluidigm <- findMarkers(fluidigm.clust,groups=fluidigm.clust$label)
markers.fluidigm
markers.fluidigm$`1`

• 如上，findMarkers的基本参数为sce对象，clust信息。

• 其中比较方法默认为t.test

  
  
  2-1

关于Top列

• 首先要知道：t.test是基于所有的clust两两之间比较的，而不是一个clust与其它所有clust的均值作比较。
• 而Top列，以1为例，表示clust1分别与另外三个clust相比差异最显著top1的基因（ranked by p-value）

chosen <- "1"  #查看clust1与其它三个clust的比较
interesting <- markers.fluidigm[[chosen]]
colnames(interesting)
interesting[1:9,1:4]

2-2

• 热图可视化

best.set <- interesting[interesting$Top <= 3,]
#选取与其它三个clust差异排名前三，共9个基因的矩阵
getMarkerEffects <- function(x, prefix="logFC", strip=TRUE, remove.na.col=FALSE) {
  #prefix="logFC"
  #x=best.set
  regex <- paste0("^", prefix, "\\.")
  i <- grep(regex, colnames(x))
  out <- as.matrix(x[,i])
  
  if (strip) {
    colnames(out) <- sub(regex, "", colnames(out))
  }
  if (remove.na.col) {
    out <- out[,!colAnyNAs(out),drop=FALSE]
  }
  
  out
}#不知为啥下载scran包没有这个函数，找到了函数源代码，同样效果
logFCs <- getMarkerEffects(best.set)
#得到行名为基因，列名为其它三类clust与clust1de1 logFC
library(pheatmap)
pheatmap(logFCs, breaks=seq(-5, 5, length.out=101))

2-3

（2）细节筛选

基于logFC

#仅挑选上调基因
markers.fluidigm.up <- findMarkers(fluidigm.clust, direction="up",
                                   groups=fluidigm.clust$label)
interesting.up <- markers.fluidigm.up[[chosen]]
interesting.up[1:10,1:4]

#logfc>1
markers.fluidigm.up2 <- findMarkers(fluidigm.clust, direction="up", lfc=1,
                                    groups=fluidigm.clust$label)
interesting.up2 <- markers.fluidigm.up2[[chosen]]
interesting.up2[1:10,1:4]

These two settings yield a more focused set of candidate marker genes that are upregulated in cluster 1

best.set <- interesting.up2[interesting.up2$Top <= 5,]
logFCs <- getMarkerEffects(best.set)
library(pheatmap)
pheatmap(logFCs, breaks=seq(-5, 5, length.out=101))

2-4

基于比较方法

• 如前所述，基础用法是所有clust两两比较；
• 可以通过设置参数pval.type=修改p.value的比较关系

（1）pval.type="all"

• A more stringent approach would only consider genes that are differentially expressed in all pairwise comparisons involving the cluster of interest.
• 即clust1的某gene必须与其它所有clust表达都有差异才认定为DE

markers.fluidigm.up3 <- findMarkers(fluidigm.clust, pval.type="all", 
                                direction="up",
                                groups=fluidigm.clust$label)
interesting.up3 <- markers.fluidigm.up3[[chosen]]
head(interesting.up3)

这种方法看起来很合理，但是也有缺点

• 此时的p.value为the maximum of the p-values from all pairwise comparisons.
• A gene will only achieve a low combined p-value if it is strongly DE in all comparisons to other clusters.
• However, any gene that is expressed at the same level in two or more clusters will simply not be detected.
• This is likely to discard many interesting genes, especially if the clusters are finely resolved with weak separation.

（2）pval.type="any"

• will be DE between at least one pair of subpopulations.
• too generous.

（3）pval.type="some"

• at least 50% of the individual pairwise comparisons exhibit no DE
• take the middle-most value as the combined p-value.
• reasonable

（3）其它参数

block=

• handle blocking factor

m.out <- findMarkers(fluidigm.clust, 
                     block=fluidigm.clust$Coverage_Type, 
                     direction="up",
                     groups=fluidigm.clust$label) 
demo <- m.out[["1"]] 
demo[demo$Top <= 5,1:4]

test=

• 默认为t.tests；
• 可以设定、采用其它方法，例如wilcox,binomial等；具体可参考原文，不再记录。

save(markers.fluidigm, file = "markergene.Rdata")

---

以上是第十一章Clustering部分的简单流程笔记，主要学习了基于t.test的marker gene detection，其它如wilcox方法见原文Chapter 10 Marker gene detection
本系列笔记基于OSCA全流程的大致流程梳理，详细原理可参考原文。如有错误，恳请指正！
此外还有刘小泽老师整理的全文翻译笔记，详见目录。