使用作者代码重复结果
课程笔记
粉丝：有单细胞线上课程吗？
小编：什么
？ 我们的单细胞转录组分析线上课程已经上线好久了，你们竟然都不知道吗，每篇推文后面的课程推荐没人看的吗，小编已哭晕在厕所
好了，戏演完了，下面郑重介绍下我们的单细胞线上课程：（详情戳下方链接）
这个课程笔记栏目记录了学员们学习单细胞转录组课程的学习笔记
希望大家能有所收获！
序言
第三单元第十二+十三讲：使用作者代码重复结果
课程链接在：http://jm.grazy.cn/index/mulitcourse/detail.html?cid=53
这一篇会是代码密集型，因为原文作者的代码真的写的很长！
下载作者的Github
1https://github.com/KPLab/SCS_CAF
文件布局如下：
下载好以后，需要将那两个tar.gz文件解压缩
看第二个R脚本 Processing.R
读入表达量数据
1# 首先指定操作路径
2Path_Main<-"~/scrna/SCS_CAF-master"
3# 然后读入原始的第一个细胞板数据
4plate1_raw<-read.delim(paste(Path_Main,"/SS2_15_0048/counts.tab",sep=""),header=TRUE,check.names=FALSE,sep="\t")
5> plate1_raw[1:3,1:3]
6 gene A3 A6
71 Adora1 0 0
82 Sntg1 0 0
93 Prim2 0 0
作者这里考虑到重复基因名的问题
1# 的确存在重复基因名
2> length(as.character(plate1_raw$gene))
3[1] 24490
4> length(unique(as.character(plate1_raw$gene)))
5> sum(duplicated(as.character(plate1_raw$gene)))
6[1] 492
7[1] 23998
8# 用这个查看
9as.character(plate1_raw$gene)[duplicated(as.character(plate1_raw$gene))]
10
11# 看一下make.unique的用法
12> make.unique(c("a", "a"))
13[1] "a" "a.1"
14# 将重复基因名变为唯一的名字
15plate1_raw$gene<-make.unique(as.character(plate1_raw$gene))
16> sum(duplicated(as.character(plate1_raw$gene)))
17[1] 0
18# 对样本重新命名
19colnames(plate1_raw)[2:length(colnames(plate1_raw))]<-paste("SS2_15_0048_",colnames(plate1_raw)[2:length(colnames(plate1_raw))],sep="")
同样的，对于0049板，也是上述操作，最后将它们按照
gene这一列进行合并，并把gene转为行名
1expr_raw<-merge(plate1_raw,plate2_raw,by="gene",all=TRUE)
2rownames(expr_raw)<-as.character(expr_raw$gene)
3expr_raw$gene<-NULL
最后计算一下dropout的比例（结果有点高）：
1# 计算dropout的比例
2sum(expr_raw==0)/(dim(expr_raw)[1]*dim(expr_raw)[2])
3# 0.8305757
读入ERCC数据
1# 也是类似上面👆的操作
2# plate1
3plate1_raw_ercc<-read.delim(paste(Path_Main,"/SS2_15_0048/counts-ercc.tab",sep=""),header=TRUE,check.names=FALSE,sep="\t")
4plate1_raw_ercc$gene<-make.unique(as.character(plate1_raw_ercc$gene))
5colnames(plate1_raw_ercc)[2:length(colnames(plate1_raw_ercc))]<-paste("SS2_15_0048_",colnames(plate1_raw_ercc)[2:length(colnames(plate1_raw_ercc))],sep="")
6
7# plate2
8plate2_raw_ercc<-read.delim(paste(Path_Main,"/SS2_15_0049/counts-ercc.tab",sep=""),header=TRUE,check.names=FALSE,sep="\t")
9plate2_raw_ercc$gene<-make.unique(as.character(plate2_raw_ercc$gene))
10colnames(plate2_raw_ercc)[2:length(colnames(plate2_raw_ercc))]<-paste("SS2_15_0049_",colnames(plate2_raw_ercc)[2:length(colnames(plate2_raw_ercc))],sep="")
11
12# 最后合并、计算ERCC dropout
13expr_raw_ercc<-merge(plate1_raw_ercc,plate2_raw_ercc,by="gene",all=TRUE)
14rownames(expr_raw_ercc)<-as.character(expr_raw_ercc$gene)
15expr_raw_ercc$gene<-NULL
16
17sum(expr_raw_ercc==0)/(dim(expr_raw_ercc)[1]*dim(expr_raw_ercc)[2])
18# 0.6267691
看一下ERCC在各个细胞的表达量分布：
1barplot(sort(as.numeric(colSums(expr_raw_ercc)),decreasing=TRUE),ylab="SPIKE LIBRARY SIZE",xlab="CELL INDEX")
然后做一个直方图，把一定数量的样本中ERCC表达量合并作一个bin：
1hist(log2(as.numeric(colSums(expr_raw_ercc))+1),col="brown",
2 main="Distribution of Spike Library Sizes",
3 xlab="Spike Library Size",breaks=20)
将内源基因与ERCC spike-in合并
先看看分别有多少：
1> print(paste0("There are ",nrow(expr_raw)," endogenous genes"))
2[1] "There are 24490 endogenous genes"
3> print(paste0("There are ",nrow(expr_raw_ercc)," spikes"))
4[1] "There are 92 spikes"
合并起来：
1all.counts.raw<-rbind(expr_raw,expr_raw_ercc)
2> dim(all.counts.raw)
3[1] 24582 768
然后重新计算dropout的比例：
1sum(all.counts.raw==0)/(dim(all.counts.raw)[1]*dim(all.counts.raw)[2])
2# 0.8298129
一共有7153个基因在所有细胞中表达量均为0：
1dim(all.counts.raw[rowSums(all.counts.raw)==0,])
2# 7153 768
关于原文去掉的52个细胞
作者也把这52个细胞的质控结果读入了R：
1cell_QC<-read.delim(paste(Path_Main,"/qc/qc_2plates.filtered_cells.txt",sep=""),row.names=1,header=TRUE,sep="\t")
2> dim(cell_QC)
3[1] 52 6
在原始矩阵中也要去掉这些细胞：
1rownames(cell_QC)<-gsub("__","_",rownames(cell_QC))
2
3all.counts.raw<-subset(all.counts.raw,select=colnames(all.counts.raw)[!colnames(all.counts.raw)%in%rownames(cell_QC)])
4> dim(all.counts.raw)
5[1] 24582 716
过滤细胞后，重新拆分成count矩阵和ERCC矩阵：
1# 得到原始count矩阵
2expr_raw<-subset(expr_raw,select=colnames(expr_raw)[!colnames(expr_raw)%in%rownames(cell_QC)])
3# 得到ERCC矩阵
4expr_raw_ercc<-subset(expr_raw_ercc,select=colnames(expr_raw_ercc)[!colnames(expr_raw_ercc)%in%rownames(cell_QC)])
分别对count矩阵和ERCC矩阵过滤
1all.counts.raw<-all.counts.raw[rowMeans(all.counts.raw)>0,]
2expr_raw<-expr_raw[rowMeans(expr_raw)>=1,]
3# count矩阵过滤后只剩下10835个基因
4expr_raw_ercc<-expr_raw_ercc[rowMeans(expr_raw_ercc)>0,]
5# ERCC也有原来的92个变成了89个
然后画CV vs Mean图
1library("matrixStats")
2# 首先还是计算CV值
3mean_expr_raw<-as.numeric(rowMeans(expr_raw,na.rm=TRUE))
4sd_expr_raw<-rowSds(as.matrix(expr_raw),na.rm=TRUE)
5cv_squared_expr_raw<-(sd_expr_raw/mean_expr_raw)^2
6# plot函数中使用(纵坐标~横坐标)
7plot(log10(cv_squared_expr_raw)~log10(mean_expr_raw),
8 pch=20,cex=0.5,xlab="log10 ( mean raw count )",
9 ylab="log10 ( CV^2)",main="RAW COUNTS")
10# 接下来添加ERCC的信息(画上红点)
11mean_expr_raw_ercc<-as.numeric(rowMeans(expr_raw_ercc,na.rm=TRUE))
12sd_expr_raw_ercc<-rowSds(as.matrix(expr_raw_ercc),na.rm=TRUE)
13cv_squared_expr_raw_ercc<-(sd_expr_raw_ercc/mean_expr_raw_ercc)^2
14points(log10(cv_squared_expr_raw_ercc)~log10(mean_expr_raw_ercc),col="red",pch=20,cex=1.5)
15# 然后对ERCC添加loess拟合曲线
16fit_expr_raw_ercc<-loess(log10(cv_squared_expr_raw_ercc)[is.finite(log10(mean_expr_raw_ercc))]~log10(mean_expr_raw_ercc)[is.finite(log10(mean_expr_raw_ercc))],span=1)
17# 从小到大排个序
18j<-order(log10(mean_expr_raw_ercc)[is.finite(log10(mean_expr_raw_ercc))])
19#
20lines(fit_expr_raw_ercc$fitted[j]~log10(mean_expr_raw_ercc)[is.finite(log10(mean_expr_raw_ercc))][j],col="red",lwd=3)
又根据拟合结果，进行了预测得到期望值，然后过滤得到符合期望CV值的基因，最后只留下5316个基因：
1pred_expr_raw<-predict(fit_expr_raw_ercc,log10(mean_expr_raw))
2filtered_expr_raw<-expr_raw[log10(cv_squared_expr_raw)>=pred_expr_raw,]
3filtered_expr_raw<-filtered_expr_raw[grepl("NA",rownames(filtered_expr_raw))==FALSE,]
4> dim(filtered_expr_raw)
5[1] 5316 716
可以看到，它的过滤从原来24490的基因，然后过滤掉没表达的基因剩10835个，然后又需要符合期望，剩5000多个。最后就是拿这5000多个基因做下游分析
看第三个R脚本 Dimensionality_reduction.R
这个脚本需要RPKM结果，因此需要先跑完上面第二个完整的脚本
降维主要使用tSNE，聚类使用dbscan（它的作用和hclust或者kmeans差不多）
上来先跑50次tSNE：
1library(Rtsne)
2N_tsne <- 50
3tsne_out <- list(length = N_tsne)
4KL <- vector(length = N_tsne)
5set.seed(1234)
6for(k in 1:N_tsne)
7{
8 tsne_out[[k]]<-Rtsne(t(log10(RPKM+1)),initial_dims=30,verbose=FALSE,check_duplicates=FALSE,
9 perplexity=27, dims=2,max_iter=5000)
10 KL[k]<-tail(tsne_out[[k]]$itercosts,1)
11 print(paste0("FINISHED ",k," TSNE ITERATION"))
12}
13names(KL) <- c(1:N_tsne)
14# 可以看到这里选择最小的KL作为50次中效果最优的tSNE，然后主要关注tsne结果的itercosts
15opt_tsne <- tsne_out[[as.numeric(names(KL)[KL==min(KL)])]]$Y
16opt_tsne_full<-tsne_out[[as.numeric(names(KL)[KL==min(KL)])]]
17save(tsne_out,opt_tsne,opt_tsne_full,file="step3-tsne-out.Rdata")
然后使用dbscan聚类：
1library(dbscan)
2plot(opt_tsne, col=dbscan(opt_tsne,eps=3.1)$cluster, pch=19, xlab="tSNE dim 1", ylab="tSNE dim 2")
如果使用kmeans方法：
1plot(opt_tsne, col=kmeans(opt_tsne,centers = 4)$clust, pch=19, xlab="tSNE dim 1", ylab="tSNE dim 2")
看看这两种聚类方法的相关性:
1> table(kmeans(opt_tsne,centers = 4)$clust,dbscan(opt_tsne,eps=3.1)$cluster)
2
3 0 1 2 3 4
4 1 1 226 0 0 0
5 2 0 0 144 0 0
6 3 0 27 0 0 44
7 4 0 231 0 43 0
8# 左侧是kmeans，上方是dbscan。
发现有一个点是离群值，所以把它放到细胞数量最多的那个组：
1library(dbscan)
2CAFgroups<-dbscan(opt_tsne,eps=3.1)$cluster
3CAFgroups_full<-dbscan(opt_tsne,eps=3.1)
4CAFgroups[CAFgroups==0]<-1
5CAFgroups_full$cluster[CAFgroups_full$cluster==0]<-1
6plot(opt_tsne, col=CAFgroups, pch=19, xlab="tSNE dim 1", ylab="tSNE dim 2")
(补充)了tSNE，还可以对PCA可视化：
1CAFgroups<-dbscan(opt_tsne,eps=3.1)$cluster
2CAFgroups_full<-dbscan(opt_tsne,eps=3.1)
3CAFgroups[CAFgroups==0]<-1
4CAFgroups_full$cluster[CAFgroups_full$cluster==0]<-1
5
6RPKM.PCA<-prcomp(log2(t(RPKM)+1), center=TRUE)
7plot(RPKM.PCA$x,main="first PCA", pch=19, col=CAFgroups)
其实有了上面这个tSNE聚类图，我们就能把基因的表达量映射上去，很像Seurat的FeaturePlot()做的那样。但是这里作者自己创造函数（参考第五个脚本：Plotting.R）
需要用到基因名、表达量矩阵、tsne坐标
1plot.feature2<-function(gene, tsne.output=tsne.out, DATAuse=DATA){
2 plot.frame<-data.frame(x=tsne.output$Y[,1], y=tsne.output$Y[,2], log2expr=as.numeric(log2(DATAuse[gene,]+1)))
3 p<-ggplot(plot.frame,aes(x=x, y=y, col=log2expr))+
4 geom_point(size=1) +
5 labs(title=paste(gene))+
6 theme_classic()+
7 scale_color_gradientn(colors = c("#FFFF00", "#FFD000","#FF0000","#360101"), limits=c(0,14))+
8 theme(axis.title = element_blank())+
9 theme(axis.text = element_blank())+
10 theme(axis.line = element_blank())+
11 theme(axis.ticks = element_blank())+
12 theme(plot.title = element_text(size=20,face="italic"))+
13 theme(legend.title = element_blank())+
14 theme(legend.position = "none")
15 return(p)
16}
17
18library(ggplot2)
19opt_tsne <- tsne_out[[as.numeric(names(KL)[KL==min(KL)])]]$Y
20opt_tsne_full<-tsne_out[[as.numeric(names(KL)[KL==min(KL)])]]
21load(file='RPKM.full.Rdata')
22load(file='CAFgroups.Rdata')
23plot.feature2("Pdgfra", opt_tsne_full, RPKM.full)
另外小提琴图的代码更是长：它是用来绘制不同基因的表达量在不同聚类分组的差异
需要用到基因名、表达量矩阵、tsne坐标
1plot.violin2 <- function(gene, DATAuse, tsne.popus, axis=FALSE, legend_position="none", gene_name=FALSE){
2 testframe<-data.frame(expression=as.numeric(DATAuse[paste(gene),]), Population=tsne.popus$cluster)
3 testframe$Population <- as.factor(testframe$Population)
4 colnames(testframe)<-c("expression", "Population")
5
6 col.mean<-vector()
7 for(i in levels(testframe$Population)){
8 col.mean<-c(col.mean,mean(testframe$expression[which(testframe$Population ==i)]))
9 }
10 col.mean<-log2(col.mean+1)
11 col.means<-vector()
12 for(i in testframe$Population){
13 col.means<-c(col.means,col.mean[as.numeric(i)])
14 }
15 testframe$Mean<-col.means
16 testframe$expression<-log2(testframe$expression+1)
17
18 p <- ggplot(testframe, aes(x=Population, y=expression, fill= Mean, color=Mean))+
19 geom_violin(scale="width") +
20 labs(title=paste(gene), y ="log2(expression)", x="Population")+
21 theme_classic() +
22
23 scale_color_gradientn(colors = c("#FFFF00", "#FFD000","#FF0000","#360101"), limits=c(0,14))+
24 scale_fill_gradientn(colors = c("#FFFF00", "#FFD000","#FF0000","#360101"), limits=c(0,14))+
25 theme(axis.title.y = element_blank())+
26 theme(axis.ticks.y = element_blank())+
27 theme(axis.line.y = element_blank())+
28 theme(axis.text.y = element_blank())+
29 theme(axis.title.x = element_blank())+
30 theme(legend.position=legend_position )
31
32 if(axis == FALSE){
33 p<-p+
34 theme( axis.line.x=element_blank(),
35 axis.text.x = element_blank(),
36 axis.ticks.x = element_blank())
37 }
38 if(gene_name == FALSE){
39 p<-p+ theme(plot.title = element_blank())
40 }else{ p<-p + theme(plot.title = element_text(size=10,face="bold"))}
41 p
42}
43
44# 例如
45plot.violin2(gene = "Pdgfra", DATAuse = RPKM.full, tsne.popus = CAFgroups_full)
看第四个R脚本 Differential_gene_expression.R
主要利用了ROTS包（Reproducibility-optimized test statistic），对每个亚群和其他几个亚群共同体进行比较
差异分析重点就在：表达矩阵和分组信息
1library(ROTS)
2library(plyr)
3# 首先针对第一亚群和其他亚群比较(把其他亚群定义为234)
4groups<-CAFgroups
5groups[groups!=1]<-234
6
7ROTS_input<-RPKM.full[rowMeans(RPKM.full)>=1,]
8ROTS_input<-as.matrix(log2(ROTS_input+1))
9# 运行代码很简单，重点就是data和group参数
10results_pop1 = ROTS(data = ROTS_input, groups = groups , B = 1000 , K = 500 , seed = 1234)
11# 最后根据FDR值得到第一组和其他组比较的差异基因
12summary_pop1<-data.frame(summary(results_pop1, fdr=1))
13head(summary_pop1)
14## Row ROTS.statistic pvalue FDR
15## Rgs5 8345 -19.89479 0 0
16## Higd1b 4559 -17.49991 0 0
17## Abcc9 393 -16.44638 0 0
18## Pdpn 7193 16.02262 0 0
19## Fbln2 3635 15.80534 0 0
20## Rgs4 8344 -15.62123 0 0
21
22# 同理，对第2组可以与1、3、4合并组比较；对第3组可以和第1、2、4组比较；对第4组可以和第1、2、3组比较
23# 都得到以后，共同保存
24save(summary_pop1,summary_pop2,summary_pop3,summary_pop4,
25 file = 'ROTS_summary_pop.Rdata')
每个亚群可以挑top18基因绘制热图
1population_subset<-c(rownames(summary_pop1[summary_pop1$ROTS.statistic<0,])[1:18],rownames(summary_pop2[summary_pop2$ROTS.statistic<0,])[1:18],rownames(summary_pop3[summary_pop3$ROTS.statistic<0,])[1:18],rownames(summary_pop4[summary_pop4$ROTS.statistic<0,])[1:18])
2RPKM_heatmap<-RPKM.full[population_subset,]
3
4RPKM_heatmap<-RPKM_heatmap[,order(CAFgroups_full$cluster)]
5RPKM_heatmap<-log2(RPKM_heatmap+1)
6
7popul.col<-sort(CAFgroups_full$cluster)
8popul.col<-replace(popul.col, popul.col==1,"#1C86EE" )
9popul.col<-replace(popul.col, popul.col==2,"#00EE00" )
10popul.col<-replace(popul.col, popul.col==3,"#FF9912" )
11popul.col<-replace(popul.col, popul.col==4,"#FF3E96" )
12library(gplots)
13
14#pdf("heatmap_genes_population.pdf")
15heatmap.2(as.matrix(RPKM_heatmap),ColSideColors = as.character(popul.col), tracecol = NA, dendrogram = "none",col=bluered, labCol = FALSE, scale="none", key = TRUE, symkey = F, symm=F, key.xlab = "", key.ylab = "", density.info = "density", key.title = "log2(RPKM+1)", keysize = 1.2, denscol="black", Colv=FALSE)
当然，原文还使用了其他几种差异分析方法，放在这里，可以做日后参考
1################################
2####### 第一种：DESeq2 ###########
3################################
4library("scran")
5library("limSolve")
6library(scater)
7library(DESeq2)
8ann<-data.frame(Plate = factor(unlist(lapply(strsplit(colnames(RPKM.full),"_"),function(x) x[3]))), Population = factor(gsub("(3|4)","2",as.character(CAFgroups)),levels=c("1","2")))
9ann<-data.frame(Population = factor(gsub("(3|4)","2",as.character(CAFgroups)),levels=c("1","2")))
10rownames(ann)<-colnames(RPKM.full)
11
12ddsFullCountTable <- DESeqDataSetFromMatrix(
13 countData = all.counts.raw[rownames(RPKM.full),],
14 colData = ann,
15 design = ~ Population)
16
17ddsFullCountTable<-DESeq(ddsFullCountTable)
18DESeq_result<-results(ddsFullCountTable)
19DESeq_result<-DESeq_result[order(DESeq_result$padj, DESeq_result$pvalue),]
20head(DESeq_result,30)
21write.table(DESeq_result[grep("ERCC", rownames(DESeq_result), invert=TRUE),], "DESeq_result.txt", col.names = TRUE, row.names = TRUE, quote = FALSE, sep="\t")
22
23################################
24####### 第二种：EdgeR ############
25################################
26library("edgeR")
27edgeR_Data<-DGEList(counts=all.counts.raw[rownames(RPKM.full),], group=ann$Population)
28edgeR_Data<-estimateCommonDisp(edgeR_Data)
29edgeR_Data<-estimateTagwiseDisp(edgeR_Data)
30edgeR_result<-exactTest(edgeR_Data)
31edgeR_result_table<-edgeR_result$table
32edgeR_result_table<-edgeR_result_table[order(edgeR_result_table$PValue),]
33edgeR_result_table<-edgeR_result_table[grep("ERCC", rownames(edgeR_result_table), invert=TRUE),]
34write.table(edgeR_result_table, "EdgeR_result.txt", col.names = TRUE, row.names = TRUE, quote=FALSE, sep="\t")
35
36################################
37###### 第三种：Wilcox ##########
38################################
39NumPerm<-1000
40POP1_expr<-subset(RPKM.full,select=rownames(ann)[ann==1])
41POP2_expr<-subset(RPKM.full,select=rownames(ann)[ann==2])
42p_wilcox<-vector()
43p_t<-vector()
44p_perm<-vector()
45statistics<-vector()
46median_POP1_expr<-vector()
47median_POP2_expr<-vector()
48a<-seq(from=0,to=length(rownames(RPKM.full)),by=1000)
49
50print("START DIFFERENTIAL GENE EXPRESSION BETWEEN POP1 AND POP2")
51for(i in 1:length(rownames(RPKM.full)))
52{
53 p_wilcox<-append(p_wilcox,wilcox.test(as.numeric(POP1_expr[rownames(RPKM.full)[i],]),as.numeric(POP2_expr[rownames(RPKM.full)[i],]))$p.value)
54 statistics<-append(statistics,wilcox.test(as.numeric(POP1_expr[rownames(RPKM.full)[i],]),as.numeric(POP2_expr[rownames(RPKM.full)[i],]))$statistic)
55 p_t<-append(p_t,t.test(as.numeric(POP1_expr[rownames(RPKM.full)[i],]),as.numeric(POP2_expr[rownames(RPKM.full)[i],]))$p.value)
56 p_perm<-append(p_perm,PermTest_Median(as.numeric(POP1_expr[rownames(RPKM.full)[i],]),as.numeric(POP2_expr[rownames(RPKM.full)[i],]),NumPerm))
57 median_POP1_expr<-append(median_POP1_expr,median(as.numeric(POP1_expr[rownames(RPKM.full)[i],])))
58 median_POP2_expr<-append(median_POP2_expr,median(as.numeric(POP2_expr[rownames(RPKM.full)[i],])))
59 if(i%in%a){print(paste("FINISHED ",i," GENES",sep=""))}
60}
61fold_change<-median_POP1_expr/median_POP2_expr
62log2_fold_change<-log2(fold_change)
63p_adj<-p.adjust(p_wilcox,method="fdr")
64output_wilcox<-data.frame(GENE=rownames(RPKM.full),POP1_EXPR=median_POP1_expr,POP234_EXPR=median_POP2_expr,FOLD_CHANGE=fold_change,LOG2FC=log2_fold_change,WILCOX_STAT=statistics,P_T_TEST=p_t,P_PERM=p_perm,P_WILCOX=p_wilcox,FDR=p_adj)
65output_wilcox<-output_wilcox[order(output_wilcox$P_PERM,output_wilcox$FDR,output_wilcox$P_WILCOX,output_wilcox$P_T_TEST,-abs(output_wilcox$LOG2FC)),]
66print(head(output_wilcox,20))
67write.table(output_wilcox,file="Wilcox_Perm_de_results.txt",col.names=TRUE,row.names=FALSE,quote=FALSE,sep="\t")
68
69################################
70####### 第四种：SCDE WORKFLOW
71################################
72library("scde")
73# factor determining cell types
74sg<-factor(as.numeric(CAFgroups))
75# the group factor should be named accordingly
76names(sg)<-colnames(expr_raw)
77table(sg)
78# define two groups of cells
79#groups<-sg
80groups <- factor(gsub("(3|4)","2",as.character(sg)),levels=c("1","2"))
81table(groups)
82# calculate models
83
84cd<-apply(expr_raw,2,function(x) {storage.mode(x) <- 'integer'; x})
85colnames(cd)<-colnames(expr_raw)
86o.ifm<-scde.error.models(counts=cd,groups=groups,n.cores=4,threshold.segmentation=TRUE,save.crossfit.plots=FALSE,save.model.plots=FALSE,verbose=1)
87print(head(o.ifm))
88# filter out cells that don't show positive correlation with
89# the expected expression magnitudes (very poor fits)
90valid.cells<-o.ifm$corr.a > 0
91table(valid.cells)
92o.ifm<-o.ifm[valid.cells, ]
93# estimate gene expression prior
94o.prior<-scde.expression.prior(models=o.ifm,counts=cd,length.out=400,show.plot=FALSE)
95# run differential expression tests on all genes.
96ediff<-scde.expression.difference(o.ifm,cd,o.prior,groups=groups,n.randomizations=100,n.cores=4,verbose=1) #batch=batch
97# top upregulated genes
98ediff_order<-ediff[order(abs(ediff$Z),decreasing=TRUE), ]
99head(ediff_order,20)
100write.table(ediff_order,file="scde_de_results_1_vs_234.txt",col.names=TRUE,row.names=TRUE,quote=FALSE,sep="\t")