Skip to content

Latest commit

 

History

History
445 lines (345 loc) · 15.2 KB

ILC_scRNA_analysis.md

File metadata and controls

445 lines (345 loc) · 15.2 KB

This is the code used for the single cell analyses of the manuscript "The heterogeneity of human CD127+ innate lymphoid cells revealed by single-cell RNA sequencing" Björklund ÅK et al.

Many of the parts that takes a long time have been set as optional where you can select to run them, or load the output files from the "data" directory. Please change the variables force.combat, force.tsne, force.pvclust and force.tsne_ilc3 to TRUE to rerun those parts of the code.

Written by Åsa K. Björklund, 2015, [email protected]

Dependencies

Cran packages:

statmod, gplots, plotrix, Rtsne, MASS, vioplot

Bioconductor:

sva

Other:

Standard pvclust package only uses spearman correlatioin, but there is an unofficial version at http://www.is.titech.ac.jp/~shimo/prog/pvclust/pvclust_unofficial_090824.zip that also includes euklidean distance.

SCDE (http://hms-dbmi.github.io/scde/)

Load some custom functions that are included in bin folder:

source("bin/functions.r")

Define color/point styles based on celltype or tonsil origin

infile = "data/ensembl_rpkmvalues_ILC.txt"
RPKM<-read.table(infile)
nS<-ncol(RPKM)
nG<-nrow(RPKM)

# define the colors for tonsil/celltype
pchdef<-1:4
tonsils<-c("T74","T75","T86")
coldef.ton<-c("black","magenta","red")
celltypes<-c("ILC1","ILC2","ILC3","NK")
coldef.cell<-c("blue","cyan","red","green")

# split names to get tonsil origin
cell<-unlist(lapply(strsplit(colnames(RPKM),"_"), function(x) x[4]))
ton<-unlist(lapply(strsplit(colnames(RPKM),"_"), function(x) x[1]))

# OBS! 6 cells from T75 P2,3 have the wrong classification by surface phenotype, change here
redef<- c("ILC1","ILC2","ILC1","ILC1","ILC1","ILC1")
redef.samples <- c("T75_P1_C6_ILC3","T75_P1_E11_ILC3","T75_P1_F4_ILC3","T75_P2_H11_ILC3","T75_P2_H2_ILC3","T75_P1_H4_ILC3")
cell[match(redef.samples,colnames(RPKM))]<-redef

# make lists for the celltypes/tonsils
sets.cell<-make_sets(cell)
sets.ton<-make_sets(ton)

# create color vector
col.ton<-make_colors(ton,tonsils,coldef.ton)
col.cell<-make_colors(cell,celltypes,coldef.cell)
pch.ton<-make_colors(ton,tonsils,pchdef)
pch.cell<-make_colors(cell,celltypes,pchdef)

# save to a file for later use
outfile<- 'data/coldef_pchdef_by_FACS.Rdata'
save(pchdef,tonsils,coldef.ton,celltypes,coldef.cell,col.ton,col.cell,pch.ton,pch.cell,sets.ton,sets.cell,file=outfile)

Find variable genes

Plot coefficient of variation vs mean and detect genes with biological variation, adapted from Brenneke et.al.

ercc<-grep("ERCC_",rownames(RPKM))


var.genes<-cv2.var.genes(RPKM[-ercc,],RPKM[ercc,],plot=T)

var.idx <- which(var.genes)
gene.names<-rownames(RPKM[-ercc,])[var.idx]
write.table(cbind(gene.names,var.idx),file="data/variable_genes_allcells.txt")

Remove batch effect

The raw data shows a clear batch effect with cells from each of the 3 tonsil donors grouping separately. Use the SVA ComBat function to remove this batch effect. OBS! Takes a long time to run, so the option of loading the file instead has been added.

library(sva)

savefile<-"data/combat_normalized.RData"
force.combat<-FALSE
if (!file.exists(savefile) || force.combat){
  cellcov<-mat.or.vec(nS,1)
  for (i in 1:4 ) {
     c<-celltypes[i]
     cellcov[sets.cell[[c]]]<-i
  }

  batchT<-as.numeric(pch.ton)
  typeT<-as.numeric(cellcov)
  pdT<-data.frame(sample=1:nS,batch=batchT,type=typeT)
  mod0<-model.matrix(~1, data=pdT)


  Rtemp<-log2(RPKM+0.1)
  cb0 = ComBat(dat=Rtemp, batch=batchT,mod=mod0,par.prior=FALSE)
  save(cb0,file=savefile)
}else{
  load(savefile) 
}

PCA before and after batch effect removal

par(mfrow=c(2,2),mar=c(4,4,2,2),xpd=T,cex.axis=0.7,cex=0.6)
# first with raw rpkm values
PC1<-pca.plot(RPKM[var.idx,],log.vals=T,main="Raw RPKM values",col=col.cell,pch=pch.ton,cex=0.6)
legend("topright",legend=celltypes,fill=coldef.cell,border=F,bty='n',cex=0.7,inset=c(0,-0.05))
legend("topleft",legend=tonsils,pch=pchdef,bty='n',cex=0.7,inset=c(0,-0.03))
pca.plot(PC1,main="Raw RPKM values",col=col.ton,pch=pch.cell,cex=0.6)
legend("topright",legend=tonsils,fill=coldef.ton,border=F,bty='n',cex=0.7,inset=c(0,-0.03))
legend("topleft",legend=celltypes,pch=pchdef,bty='n',cex=0.7,inset=c(0,-0.05))

# then after ComBat normalization with rpkm values
PC2<-pca.plot(cb0[var.idx,],log.vals=F,main="ComBat normalized",col=col.cell,pch=pch.ton,cex=0.6)
legend("topright",legend=celltypes,fill=coldef.cell,border=F,bty='n',cex=0.7,inset=c(0,-0.05))
legend("topleft",legend=tonsils,pch=pchdef,bty='n',cex=0.7,inset=c(0,-0.03))
pca.plot(PC2,main="ComBat normalized",col=col.ton,pch=pch.cell,cex=0.6)
legend("topright",legend=tonsils,fill=coldef.ton,border=F,bty='n',cex=0.7,inset=c(0,-0.03))
legend("topleft",legend=celltypes,pch=pchdef,bty='n',cex=0.7,inset=c(0,-0.05))

Run tSNE

Run tSNE (Barnes-Hut tSNE in Rtsne package), using 10 first principal components as input. Tsne was run for 20 iterations since the results are slightly differen each time. The results are stored in a list RES.

library(Rtsne)

savefile1 <- "data/tsne_10Pc_20i.Rdata"
force.tsne<-FALSE
if (!file.exists(savefile1) || force.tsne){
  C<-cb0[var.genes[,2],]
  RES<-list()
  pdf("figures/tsne_10PC_20i.pdf")
  par(mfrow=c(4,4),mar=c(1,1,3,1))
  for (i in 1:20){
       tsne.out10<- Rtsne(t(C),initial_dims=10,dims=2,theta=0.001)
       plot(tsne.out10$Y,col=col.cell,pch=pch.ton,ylab='',xlab='')
       RES[[i]]<-tsne.out10
  }
  dev.off()
  save(RES, file=savefile1)
}else{
  load(savefile1)   
}

plot(RES[[1]]$Y,col=col.cell,pch=pch.ton,ylab='',xlab='')

Bootstrapped clustering with pvclust

pvclust based on euclidean distances instead of correlations are done using modified pvclust package from: http://www.is.titech.ac.jp/~shimo/prog/pvclust/pvclust_unofficial_090824.zip

# set to whaterver path you have for the pvclust package:
source("/Users/asab/jobb/data/local/bin/pvclust_unofficial_090824/pvclust.R")
source("/Users/asab/jobb/data/local/bin/pvclust_unofficial_090824/pvclust-internal.R")
library(MASS)

Rsum<-Reduce("cbind",lapply(RES,function(x) x$Y))

savefile2<-"data/pvclust_euclidean_10PC_20i.Rdata"
force.pvclust<-FALSE
if (!file.exists(savefile2) || force.pvclust){
   bhc<-pvclust(t(Rsum), method.hclust="ward",method.dist="euclidean")
   save(bhc,file=savefile2)
}else {
   load(savefile2)
}

Define cluster groups

groups<-cutree(bhc$hclust,4)
group.def<-c("ILC1","NK","ILC2","ILC3")
cluster.groups<-mat.or.vec(1,nS)
for (i in 1:4) {
    cluster.groups[groups==i]<-group.def[i]
}

write.table(cbind(cluster.groups[1,],colnames(RPKM)),file="data/cluster_groups_pvclust.txt")

sets.cl<-make_sets(cluster.groups)
col.cl<-make_colors(cluster.groups,celltypes,coldef.cell)

Plot dendrogram

plot(bhc,main="pvclust, euclidean distance",cex.pv=1.2)

Plot heatmap with hallmark genes

Plot the genes in data/hallmark_gene_list.txt as a heatmap with cells ordered by the pvclust dendrogram.

# define colors:
cc<-colorRampPalette(c("yellow","red","black"))
cols<-cc(100)
yrange<-c(0,14)
colbr <- seq(yrange[1], yrange[2], len=101)

# read in gene list
G<-read.table("data/hallmark_gene_list.txt",sep="\t",header=T)
gene.col<-make_colors(G[,1],celltypes,coldef.cell)
gene.idx<-match(G[,3],rownames(RPKM))

dendro<-as.dendrogram(bhc$hclust)
library(gplots)
h <- heatmap.2(log2(as.matrix(RPKM[gene.idx,])+1),col=cols,breaks=colbr,scale="none",
  trace="none",cexRow=0.5,cexCol=0.5,ColSideColors=col.cell,RowSideColors=gene.col,
  key.title=NA,key.xlab="log2(rpkm)", key.ylab=NA,Colv=dendro,Rowv=FALSE)

Plot violin plots

This example code snippet takes the gene list data/hallmark_gene_list.txt, but can be applied to any list of genes, thus producing the violin plots in Figures 4-7 of the manuscript.

crange.log<-seq(0,14,by=0.1)
col.log<-cc(length(crange.log))
R<-as.matrix(RPKM)

par(mfcol=c(20,5),xpd=T,cex=0.5)
par(mar=c(0,7,0,0))
for (i in 1:nrow(G)){
    l<-lapply(sets.cl,function(x) R[gene.idx[i],x] )
    gname<-G[i,3]
    make.vioplot(l,name=gname,text.col=NA,ylim=yrange,horizontal=F)
}
plot.color.bar()

Protein and RNA comparison

All marker protein index data and log2(rpkm) values for the corresponding mRNAs can be found in file data/protein_RNA_data_markers.txt

D<-read.table('data/protein_RNA_data_markers.txt')

# remove negative markers, NKG2A, CD16, FSC and SSC
rem<-c(4,5,13,14,19,20)
D<-D[,-rem]

nS<-nrow(D)
nD<-7
prot.idx<-1:7
rna.idx<-8:14
markers<-colnames(D)[prot.idx]
genes<-colnames(D)[rna.idx]

tonsils<-c("T75","T74","T86")
ton2 <-unlist(lapply(strsplit(rownames(D),"_"), function(x) x[1]))
sets.ton2<-make_sets(ton2)

# normalize data for each sort occation

normdata<-mat.or.vec(nS,nD*2)
for (t in names(sets.ton2)){
    d<-sets.ton2[[t]]
    normdata[d,]<-apply(D[d,],2,norm_distr)
}

# plot
par(mfrow=c(2,2),cex=0.5,mar=c(5,5,2,2))
# first nkp44, first column
i<-1
data<-normdata
c<-cor(data[,i],data[,i+nD])
cs<-cor(data[,i],data[,i+nD],method="spearman")
smoothScatter(data[,i],data[,i+nD],xlab=sprintf("%s RNA abundance", genes[i]), ylab=sprintf("%s Protein abundance",markers[i]),main=sprintf("%s vs %s, Pearson=%.4f\nSpearman=%.4f",genes[i],markers[i],c,cs),cex.main=0.8)

# then mean all
plot_mean(normdata,"All")

# then simplots with spearman, pearson
run_simulation_plot(normdata,1000,noPar=TRUE)

Clustering of ILC3 cells

All cells defined as ILC3 cells were clustered in a similar manner to the whole dataset, using several iterations of tSNE and hierarchical clustering, but the gene set used consisted of immune related genes only.

immune.data<-read.table("data/GO_immune_genes.txt")
ilc3s<-sets.cl$ILC3
c.temp<-cb0[immune.data[,2],ilc3s]

library(Rtsne)

savefile.ilc3<-"data/tsne_ilc3_50i.Rdata"
force.ilc3_tsne<-FALSE
if (!file.exists(savefile.ilc3) || force.ilc3_tsne){
  RES3<-list()
  for (i in 1:50){
    tsne.out<- Rtsne(t(c.temp),initial_dims=3,dims=2,perplexity=20,theta=0.1)
    RES3[[i]]<-tsne.out
  }
  save(RES3,file=savefile.ilc3)
}else {
  load(savefile.ilc3)
}


Rsum<-Reduce("cbind",lapply(RES3,function(x) x$Y))
hcl<-hclust(dist(Rsum),method="ward.D2")
groups<-cutree(hcl,3)
sets.ilc3<-make_sets(groups)
coldef.ilc3<-c("red","blue","orange")
col.ilc3<-make_colors(groups,1:3,coldef.ilc3)

PCA with ILC3 cells

All cells defined as ILC3 cells were clustered in a similar manner to the whole dataset, using several iterations of tSNE and hierarchical clustering.

PC<-pca.plot(c.temp,log.vals=F,col=col.ilc3,pch=16)

# plot gene loadings onto plot
loading.genes<- c("NCR2","TYROBP","KIT","CD3E","SELL","IL1R1","TNFSF10","PRAM1","HLA-DRA","HLA-DRB1","HLA-DRB5","HLA-DPA1","HLA-DPB1")
loading.idx<-match(loading.genes,rownames(c.temp))
data<-PC$x[,c(1,2)]
rot<-PC$rotation[,c(1,2)]
#multiplication factor
mult <- min(
  (max(data[,2]) - min(data[,2])/(max(rot[,2])-min(rot[,2]))),
  (max(data[,1]) - min(data[,1])/(max(rot[,1])-min(rot[,1])))
)
v1<-rot[,1]*mult
v2<-rot[,2]*mult
text(v1[loading.idx],v2[loading.idx],rownames(c.temp)[loading.idx],cex=0.7,col="black")

tSNE plots for ICL3s colored by different markers

ilc3.genes<-c("NCR2", "KIT", "CD3E", "SELL", "HLA-DRA", "HLA-DRB1", "HLA-DPA1", "IL1R1" )
par(mfrow=c(4,5),mar=c(1,1,3,1))
tsne.out<-RES3[[i]]
plot(tsne.out$Y,col=col.ilc3,pch=16,main="Clusters",axes=F,xlab=F,ylab=F,cex=0.7)
plot(tsne.out$Y,col=col.ton[ilc3s],pch=16,main="Donors",axes=F,xlab=F,ylab=F,cex=0.7)
for (g in ilc3.genes){
  g.idx<-match(g,rownames(R))
  col.g<-convert.to.color(log2(R[g.idx,ilc3s]+1),cc,col.range=crange.log) 
  plot(tsne.out$Y,col=col.g$cols,main=g,axes=F,xlab=F,ylab=F,pch=16,cex=0.7)   
}
plot.color.bar()

# detected genes
nDet<-apply(R[,ilc3s],2,function(x) length(which(x>1)))
cols.nDet<-convert.to.color(nDet,cc)
plot(tsne.out$Y,col=cols.nDet$cols,main="Detected genes",axes=F,xlab=F,ylab=F,pch=16,cex=0.7)
plot.color.bar(col=cols.nDet$col.def,crange=cols.nDet$col.range,main="Detected genes")

# NKp44
D<-read.table('data/protein_RNA_data_markers.txt')
nkp44<-D[rownames(D) %in% colnames(R)[ilc3s],1]
cols.nkp44<-convert.to.color(nkp44,cc)
plot(tsne.out$Y,col=cols.nkp44$cols,main="nKp44 intensity",axes=F,xlab=F,ylab=F,pch=16,cex=0.7)
plot.color.bar(col=cols.nkp44$col.def,crange=cols.nkp44$col.range,main="nKp44 intensity")

# Forward scattering
cols.fsc <- convert.to.color(D[rownames(D) %in% colnames(R)[ilc3s],19],cc)
plot(tsne.out$Y,col=cols.fsc$cols,main="FSC",axes=F,xlab=F,ylab=F,pch=16,cex=0.7)
plot.color.bar(col=cols.fsc$col.def,crange=cols.fsc$col.range,main="FSC")

Finally, the session info:

sessionInfo()

## R version 3.2.2 (2015-08-14)
## Platform: x86_64-apple-darwin13.4.0 (64-bit)
## Running under: OS X 10.10.5 (Yosemite)
## 
## locale:
## [1] C
## 
## attached base packages:
## [1] stats     graphics  grDevices utils     datasets  methods   base     
## 
## other attached packages:
##  [1] vioplot_0.2       sm_2.2-5.4        gplots_2.17.0    
##  [4] MASS_7.3-45       Rtsne_0.10        sva_3.14.0       
##  [7] genefilter_1.50.0 mgcv_1.8-9        nlme_3.1-122     
## [10] statmod_1.4.22   
## 
## loaded via a namespace (and not attached):
##  [1] Rcpp_0.12.2          formatR_1.2.1        GenomeInfoDb_1.4.3  
##  [4] bitops_1.0-6         tools_3.2.2          digest_0.6.8        
##  [7] annotate_1.46.1      evaluate_0.8         RSQLite_1.0.0       
## [10] lattice_0.20-33      Matrix_1.2-2         DBI_0.3.1           
## [13] yaml_2.1.13          parallel_3.2.2       stringr_1.0.0       
## [16] knitr_1.11           caTools_1.17.1       gtools_3.5.0        
## [19] S4Vectors_0.6.6      IRanges_2.2.9        stats4_3.2.2        
## [22] grid_3.2.2           Biobase_2.28.0       AnnotationDbi_1.30.1
## [25] XML_3.98-1.3         survival_2.38-3      rmarkdown_0.8.1     
## [28] gdata_2.17.0         magrittr_1.5         htmltools_0.2.6     
## [31] BiocGenerics_0.14.0  splines_3.2.2        xtable_1.8-0        
## [34] KernSmooth_2.23-15   stringi_1.0-1