#determine which test method to use method = as.character(commandArgs()[4]); if (method == "glm") { ### This program is for Generalized Linear Model (GLM) analysis ### Y = X + covar1 + covar2 ...... ### Usually (not limited), ### Y is clinical trait (quantitative or binary) ### X is variant/gene/mutation etc. ### covar means covariate ### by Qunyuan Zhang (qunyuan@wustl.edu), 02/16/2012 updated ### input options model.file = as.character(commandArgs()[5]); y.file = as.character(commandArgs()[6]); x.file = as.character(commandArgs()[7]); out.file = as.character(commandArgs()[8]); x.names="*"; ### to run it on command line, type # R --no-save < glm.R model.file y.file x.file * out.csv # or (if you want to define x variables) # R --no-save < glm.R model.csv y.file x.file x1|x2|x3 out.csv # or (if x.file has been merged into y.file) # R --no-save < glm.R model.csv y.file * x1|x2|x3 out.csv # or (if you have defined x variable in model.file) # R --no-save < glm.R model.csv y.file * * out.csv ### about model.file # column "type": Q=quantitative trait, B=binary trait # column "y": trait name # colum "x": variant/gene name; if x=NA or blank, it will determined by x.file and x.names # column "cvar": covariate(s) # tab delimited ### about y.file # trait data file, column 1 must be sample id # tab delimited ### about x.file # usually mutation/variant/geene data file # the first column must be sample id (the same as in y.file, ordered or not) # tab delimited # x.file="*" if x.file already merged into y.file ### about x.names # x.names="*" will use all column names in x.file as x variable names # or you can define it in the format x.names="gene1|gene2|gene3" # self-defined x.names have to be found in column names of y.file and/or x.file #################### myglm fuction ############## myglm=function(z,trait,variant,covar=NA,ytype) { if (nchar(covar)==0 | is.na(covar) | is.null(covar)) { model=formula(paste(trait,"~",variant)) } else { model=formula(paste(trait,"~",variant,"+",covar)) } if (ytype=="B") fit=glm(formula=model,data=z,family=binomial(link = "logit")) if (ytype=="Q") fit=glm(formula=model,data=z,family=gaussian(link = "identity")) fit } ################################################# ### data input ##### read.table(model.file,colClasses="character",na.strings = c("","NA"),sep="\t",header=T)->md read.table(y.file,na.strings = c("","NA"),sep="\t",header=T)->y if (x.names!="*") x.names=strsplit(x.names,split="[|]")[[1]] if (x.file!="*") { read.table(x.file,na.strings = c("","NA"),sep="\t",header=T)->x xid=colnames(x)[1] xs=colnames(x)[-1] if (x.names!="*") {x=x[,c(xid,x.names)];xs=colnames(x)[-1]} x.names=xs yid=colnames(y)[1] ysid = ! (colnames(y) %in% xs) y=y[,ysid] if (sum(ysid)==1) {y=data.frame(id=y);colnames(y)[1]=yid} #y=merge(y,x,by.x = xid, by.y = yid) y=merge(x,y,by.x = xid, by.y = yid) } ######### analysis ########## tt=NULL for (i in c(1:nrow(md))) { ytype=md[i,1];yi=md[i,2];xs=md[i,3];covi=md[i,4];memo=md[i,5] if (!is.na(xs) & nchar(xs)>0) xs=strsplit(xs,split="[|]")[[1]] if (is.na(xs)[1]|nchar(xs)[1]==0) xs=x.names if (length(covi)==0) covi=NA for (xi in xs) { print(yi); print(xi); print(covi); print("******") if (ytype=="Q") try(anova(myglm(y,yi,xi,covi,ytype),test="F"))->fit if (ytype=="B") try(anova(myglm(y,yi,xi,covi,ytype),test="Chisq"))->fit if (class(fit)[1]!="try-error") { fit=as.matrix(fit) if (xi %in% rownames(fit)) tt=rbind(tt, cbind(yi,ytype,xi,as.data.frame(t(fit[xi,])),covi,memo)) } } } #"yi","ytype","xi","Df","Deviance","Resid. Df","Resid. Dev","F","Pr(>F)","covi","memo" if (ytype=="Q") colnames(tt) = c("y","y_type","x","degrees_freedom","deviance","residual_degrees_freedom","residual_deviance","F_statistic","p-value","covariants","memo"); if (ytype=="B") colnames(tt) = c("y","y_type","x","degrees_freedom","deviance","residual_degrees_freedom","residual_deviance","p-value","covariants","memo"); write.table(tt,file=out.file,quote=F,sep="\t",row.names=F); } else { # else, we process numerical or categorical clinical data correlation clinical_data = as.character(commandArgs()[5]); mutation_matrix = as.character(commandArgs()[6]); output_file = as.character(commandArgs()[7]); # FUNCTION finds the correlation between two variables cor2=function(ty,tx,method) { id=intersect(!is.na(ty),!is.na(tx)); ty=ty[id]; tx=tx[id]; if(method=="cor") { tst=cor.test(tx,ty); s=tst$est; p=tst$p.value; } if(method=="wilcox") #x must be (0,1) mutation data { tst=wilcox.test(x=ty[tx==0],y=ty[tx>=1]) s=tst$stat p=tst$p.value } if(method=="chisq") { tst=chisq.test(tx,ty); s=tst$stat; p=tst$p.value; } if(method=="fisher") { tst=fisher.test(tx,ty) s=tst$p.value p=tst$p.value } if(method=="anova") { tst=summary(aov(ty~tx,as.data.frame(cbind(tx,ty)))) s=tst[[1]]$F[1] p=tst[[1]]$Pr[1] } tt=c(p,s); tt; } # END cor2 # FUNCTION runs correlation test on matrixes of data cor2test =function(y,x=NULL,method="cor",cutoff=1,sep="\t",outf=NULL) { if (!is.null(x)) { if (length(x)==1) {read.table(x,header=T,sep=sep)->x;} if (length(y)==1) {read.table(y,header=T,sep=sep)->y;} colnames(y)[1]="id"; colnames(x)[1]="id"; tt=character(0); for (vi in colnames(x)[-1]) { for (vj in colnames(y)[-1]) { tx=x[,c("id",vi)]; tx=tx[!is.na(tx[,vi]),]; tx=tx[!duplicated(tx[,"id"]),]; ty=y[,c("id",vj)]; ty=ty[!is.na(ty[,vj]),]; ty=ty[!duplicated(ty[,"id"]),]; xy=merge(tx,ty,by.x="id",by.y="id"); tx=xy[,2]; ty=xy[,3]; n=length(xy[,"id"]); rst=try(cor2(ty,tx,method)); if (class(rst)=="try-error") {p=NA;s=NA;} else {p=rst[1];s=rst[2];} t=c(vi,vj,method,n,s,p) tt=rbind(tt,t); } #end vj } #end vi rownames(tt)=NULL; colnames(tt)=c("x","y","method","n","s","p"); tt=as.data.frame(tt); tt[,"s"]=as.character(tt[,"s"]); tt[,"s"]=as.numeric(tt[,"s"]); tt[,"p"]=as.character(tt[,"p"]); tt[,"p"]=as.numeric(tt[,"p"]); fdr=p.adjust(tt[,"p"],method="fdr"); bon=p.adjust(tt[,"p"],method="bon"); tt=cbind(tt,fdr,bon); tt=tt[order(tt[,"p"]),]; } if (is.null(x)) { if (length(y)==1) {read.table(y,header=T,sep=sep)->y;} x=y; nxy=ncol(y)-1; colnames(y)[1]="id"; colnames(x)[1]="id" tt=character(0); for (i in c(1:(nxy-1))) { for (j in c((i+1):nxy)) { vi=colnames(x)[-1][i]; vj=colnames(y)[-1][j]; tx=x[,c("id",vi)]; tx=tx[!is.na(tx[,vi]),]; tx=tx[!duplicated(tx[,"id"]),] ty=y[,c("id",vj)]; ty=ty[!is.na(ty[,vj]),]; ty=ty[!duplicated(ty[,"id"]),]; xy=merge(tx,ty,by.x="id",by.y="id"); tx=xy[,2]; ty=xy[,3]; n=length(xy[,"id"]); rst=try(cor2(ty,tx,method)); if (class(rst)=="try-error") {p=NA;s=NA;} else {p=rst[1];s=rst[2];} t=c(vi,vj,method,n,s,p); tt=rbind(tt,t); } #end vj } #end vi rownames(tt)=NULL; colnames(tt)=c("x","y","method","n","s","p"); tt=as.data.frame(tt); tt[,"s"]=as.character(tt[,"s"]); tt[,"s"]=as.numeric(tt[,"s"]); tt[,"p"]=as.character(tt[,"p"]); tt[,"p"]=as.numeric(tt[,"p"]); fdr=p.adjust(tt[,"p"],method="fdr"); bon=p.adjust(tt[,"p"],method="bon"); tt=cbind(tt,fdr,bon); tt=tt[order(tt[,"p"]),]; } if (!is.null(outf)) { colnames(tt)=c("x","y","method","n","s","p","fdr","bon"); #The amount of precision that R prints with is somehow machine dependent (or the R version?) tt[,"s"] = sapply(tt[,"s"], sprintf, fmt="%.4E"); tt[,"p"] = sapply(tt[,"p"], sprintf, fmt="%.4E"); tt[,"fdr"] = sapply(tt[,"fdr"], sprintf, fmt="%.2E"); tt[,"bon"] = sapply(tt[,"bon"], sprintf, fmt="%.2E"); #The ordering should be done after reformatting the precision (duh) tt=tt[order(tt[,"x"]),]; tt=tt[order(tt[,"p"]),]; write.table(tt,file=outf,quote=FALSE,row.names=FALSE,sep=","); } invisible(tt); } #END cor2test #run correlation test using function cor2test(y = clinical_data, x = mutation_matrix, method = method, outf = output_file); }