--- title: Sparse regression with paired covariates header-includes: \usepackage{amsmath} \usepackage{xcolor} \usepackage{caption}\captionsetup[figure]{labelformat=empty} \usepackage[section]{placeins} \definecolor{blue}{HTML}{0000CD} \definecolor{red}{HTML}{CD0000} output: rmarkdown::html_vignette vignette: > %\VignetteIndexEntry{script} %\VignetteEncoding{UTF-8} %\VignetteEngine{knitr::rmarkdown} editor_options: chunk_output_type: console --- Code for reproducing the results shown in the manuscript. ```{r setup,include=FALSE} # Set eval to TRUE to create Figures. # Set echo to TRUE for html, but to FALSE for pdf. knitr::opts_chunk$set(eval=FALSE,echo=TRUE,fig.path="images/") eval <- FALSE # TRUE or FALSE ``` # Initialisation ```{r functions} ### Loading functions. ### inst <- rownames(utils::installed.packages()) cran <- c("devtools","R.utils","Matrix","glmnet","pROC","BiocManager","ashr") # "googledrive", "httpuv" if(!all(cran %in% inst)){ for(i in seq_along(cran)){ if(!cran[i] %in% inst){ install.packages(cran[i]) } } } bioc <- c("edgeR","TCGAbiolinks") if(!all(bioc %in% inst)){ #source("http://bioconductor.org/biocLite.R") for(i in seq_along(bioc)){ if(!bioc[i] %in% inst){ #biocLite(bioc[i]) BiocManager::install(bioc[i]) } } } #if(!"ashr" %in% inst){ # devtools::install_github("stephens999/ashr") #} user <- Sys.getenv("USERNAME") path <- file.path("C:","Users",user,"Desktop","palasso") if(user=="arra"){path <- "C:/Users/arra/Desktop/MATHS/palasso_desktop"} if(user==""){path <- "/virdir/Scratch/arauschenberger/palasso"} setwd(path) folders <- c("data","results") invisible(sapply(folders,function(x) if(!dir.exists(x)){dir.create(x)})) if(user!="arra"){ devtools::install_github("kkdey/CorShrink") # ref="a9f6ba0" devtools::install_github("rauschenberger/palasso") # ref="4a995a2" } if(FALSE){ # The functions <>, <> and <> access the hard disk, but also try to access googledrive. save <- function(object,file){ base::save(object,file=file) tryCatch(expr=googledrive::drive_upload(media=file,path=file), error=function(x) NULL) #Sys.sleep(0.5) } file.exists <- function(file){ offline <- base::file.exists(file) online <- FALSE if(!offline){ d <- googledrive::as_dribble(x=file) online <- tryCatch(expr=googledrive::some_files(d), error=function(x) FALSE) #Sys.sleep(0.5) } return(offline|online) } file.remove <- function(file){ base::file.remove(file) tryCatch(expr=googledrive::drive_trash(file=file), error=function(x) NULL) #Sys.sleep(0.5) } # The function <> moves files from the research cloud to the remote drive. Given both paths, it first verifies whether the folders SIM, GDC and CCA are available, and then copies all missing files from the research cloud to the remote drive. # from: path to the origin # to: path to the destination update <- function(from,to){ dir <- c("SIM","GDC","CCA") if(any(!dir.exists(file.path(from,dir)))){stop("Invalid.")} if(any(!dir.exists(file.path(to,dir)))){stop("Invalid.")} pb <- utils::txtProgressBar(min=0,max=1,style=3) for(i in seq_along(dir)){ files0 <- dir(file.path(from,dir[i])) files1 <- dir(file.path(to,dir[i])) names <- files0[!files0 %in% files1] for(j in seq_along(names)){ utils::setTxtProgressBar(pb=pb,value=(i-1)/3+(j*i)/(3*length(names))) file.copy(from=file.path(from,dir[i],names[j]), to=file.path(to,dir[i],names[j]), copy.date=TRUE) } utils::setTxtProgressBar(pb=pb,value=i/3) } } # update(from="results",to="//tsclient/N/palasso/results") } ``` # Preparation Last data download started on 2019-03-26 (R version 3.5.3). ```{r get_isoform,eval=FALSE} ### Downloading "Isoform Expression Quantification". ### #rm(list=ls()) #<> directory <- file.path(path,"data") setwd(directory) # Retrieving cancer types: project <- TCGAbiolinks::getGDCprojects()$id project <- project[grepl(x=project,pattern="TCGA")] # Downloading isoform expression data: y <- X <- list() for(i in seq_along(project)){ query <- TCGAbiolinks::GDCquery(project=project[i], data.category="Transcriptome Profiling", data.type="Isoform Expression Quantification") TCGAbiolinks::GDCdownload(query,method="api",directory=directory) trace(TCGAbiolinks:::readTranscriptomeProfiling,tracer=quote(ignore.case<-TRUE)) X[[i]] <- TCGAbiolinks::GDCprepare(query,directory=directory) X[[i]][,c("miRNA_ID","reads_per_million_miRNA_mapped", "cross-mapped","miRNA_region")] <- NULL y[[i]] <- rep(project[i],times=length(unique(X[[i]]$barcode))) } save(list=c("y","X"),file=file.path(path,"data","isoform_raw.RData")) load(file.path(path,"data","isoform_raw.RData"),verbose=TRUE) # Merging isoform expression data: Xs <- do.call(what=rbind,args=X) # sparse matrix y <- do.call(what="c",args=y) # Transform to matrix Xs$isoform_coords <- gsub(pattern="hg38:chr",replacement="",x=Xs$isoform_coords) samples <- unique(Xs$barcode) covariates <- unique(Xs$isoform_coords) row <- match(Xs$barcode,samples) col <- match(Xs$isoform_coords,covariates) X <- Matrix::sparseMatrix(i=row,j=col,x=Xs$read_count,dimnames=list(samples,covariates)) # Order by molecular location split <- strsplit(x=colnames(X),split=":|-") chr <- sapply(split,function(x) x[[1]]) pos <- sapply(split,function(x) x[[2]]) order <- order(chr,pos) X <- X[,order] if(FALSE){ # testing i <- sample(seq_len(nrow(Xs)),size=1) Xs$read_count[i] X[Xs$barcode[i],Xs$isoform_coords[i]] } save(list=c("y","X"),file=file.path(path,"data","isoform_all.RData")) ``` ```{r get_miRNA,eval=FALSE} ### Downloading "miRNA Expression Quantification". ### #rm(list=ls()) #<> directory <- file.path(path,"data") setwd(directory) # Downloading data. project <- TCGAbiolinks::getGDCprojects()$id project <- project[grepl(x=project,pattern="TCGA")] y <- X <- list() for(i in seq_along(project)){ query <- TCGAbiolinks::GDCquery(project=project[i], data.category="Transcriptome Profiling", data.type="miRNA Expression Quantification") TCGAbiolinks::GDCdownload(query,method="api",directory=directory) trace(TCGAbiolinks:::readTranscriptomeProfiling,tracer=quote(ignore.case<-TRUE)) data <- TCGAbiolinks::GDCprepare(query,directory=directory) X[[i]] <- t(data[,c(seq(from=2,to=ncol(data),by=3))]) y[[i]] <- rep(project[i],times=nrow(X[[i]])) } save(list=c("y","X"),file=file.path(path,"data","miRNA_raw.RData")) load(file.path(path,"data","miRNA_raw.RData")) X <- do.call(what=rbind,args=X) y <- do.call(what="c",args=y) rownames(X) <- gsub(pattern="read_count_",replacement="",x=rownames(X)) save(list=c("y","X"),file=file.path(path,"data","miRNA_all.RData")) ``` ```{r get_gene,eval=FALSE} ### Downloading "Gene Expression Quantification". ### #rm(list=ls()) #<> directory <- file.path(path,"data") setwd(directory) # Retrieving cancer types: project <- TCGAbiolinks::getGDCprojects()$id project <- project[grepl(x=project,pattern="TCGA")] # Downloading data: memory.limit(size=16000) # Activate virtual memory in system control! y <- X <- list() for(i in seq_along(project)){ query <- TCGAbiolinks::GDCquery(project=project[i], data.category="Transcriptome Profiling", data.type="Gene Expression Quantification", workflow.type="HTSeq - Counts"); gc() TCGAbiolinks::GDCdownload(query=query,method="api",directory=directory); gc() trace(TCGAbiolinks:::readTranscriptomeProfiling,tracer=quote(ignore.case<-TRUE)); gc() X[[i]] <- TCGAbiolinks::GDCprepare(query,directory=directory); gc() y[[i]] <- rep(project[i],times=ncol(X[[i]])); gc() } save(list=c("y","X"),file=file.path(path,"data","gene_raw.RData")) load(file.path(path,"data","gene_raw.RData")) genes <- SummarizedExperiment::rowData(X[[1]]) mart <- biomaRt::useMart("ensembl",dataset="hsapiens_gene_ensembl") # char <- biomaRt::getBM(attributes=c("ensembl_gene_id","chromosome_name","transcript_start","gene_biotype"),filters=c("biotype","chromosome_name"),values=list("protein_coding",c(1:22,"X")),mart=mart) select <- genes$ensembl_gene_id[genes$ensembl_gene_id %in% char$ensembl_gene_id] X <- lapply(X,function(x) t(SummarizedExperiment::assays(x)$"HTSeq - Counts"[select,])) X <- do.call(what=rbind,args=X) y <- do.call(what="c",args=y) save(list=c("y","X"),file=file.path(path,"data","gene_all.RData")) ``` ```{r get_CNV,eval=FALSE} ### Downloading "Copy Number Variation". ### #rm(list=ls()) #<> directory <- file.path(path,"data") setwd(directory) project <- TCGAbiolinks::getGDCprojects()$id project <- project[grepl(x=project,pattern="TCGA")] y <- X <- list() for(i in seq_along(project)){ query <- TCGAbiolinks::GDCquery(project=project[i], data.category="Copy Number Variation", data.type="Masked Copy Number Segment") TCGAbiolinks::GDCdownload(query=query,method="api",directory=directory) trace(TCGAbiolinks:::readTranscriptomeProfiling,tracer=quote(ignore.case<-TRUE)) X[[i]] <- TCGAbiolinks::GDCprepare(query,directory=directory) y[[i]] <- rep(project[i],times=length(unique(X[[i]]$Sample))) # correct? } save(list=c("y","X"),file=file.path(path,"data","CNV_raw.RData")) load(file.path(path,"data","CNV_raw.RData"),verbose=TRUE) # Merging CNV data: Xs <- do.call(what=rbind,args=X) # sparse matrix y <- do.call(what="c",args=y) #table(Xs$Sample) # Prepare cutting. cut <- list() cut$chr <- c(1:22,"X") cut$start <- sapply(cut$chr,function(x) min(Xs$Start[Xs$Chromosome==x])) cut$end <- sapply(cut$chr,function(x) max(Xs$End[Xs$Chromosome==x])) cut$length <- cut$end-cut$start cut$dist <- sum(cut$length)/10000 cut$num <- round(cut$length/cut$dist) # Create covariates. cov <- list() cov$p <- sum(cut$num) cov$chromosome <- unlist(sapply(cut$chr,function(i) rep(i,times=cut$num[i]))) cov$location <- unlist(sapply(cut$chr,function(i) round(seq(from=cut$start[i],to=cut$end[i],length.out=cut$num[i])))) cov$name <- paste0(cov$chromosome,":",cov$location) # Create indices for each covariate. index <- rep(list(integer()),times=cov$p) pb <- utils::txtProgressBar(min=0,max=cov$p,style=3) for(j in seq_len(cov$p)){ utils::setTxtProgressBar(pb=pb,value=j) index[[j]] <- which((Xs$Chromosome==cov$chromosome[j]) & (Xs$Start<=cov$location[j]) & (cov$location[j]<=Xs$End)) # consider < } # Expand indices to matrix. X <- matrix(0,nrow=length(unique(Xs$Sample)),ncol=cov$p, dimnames=list(unique(Xs$Sample),cov$name)) for(j in seq_along(index)){ mean <- Xs$Segment_Mean[index[[j]]] i <- Xs$Sample[index[[j]]] X[i,j] <- mean } if(FALSE){ # test sample <- sample(rownames(X),size=1) covariate <- sample(colnames(X),size=1) split <- strsplit(covariate,split=":")[[1]] a <- X[sample,covariate] b <- Xs$Segment_Mean[(Xs$Sample==sample) & (Xs$Chromosome==split[1]) & (Xs$Start<=as.numeric(split[2])) & (as.numeric(split[2]) < Xs$End)] all(a==b) } save(list=c("y","X","index"),file=file.path(path,"data","CNV_all.RData")) ``` ```{r do_filter,eval=FALSE} ### Extracting samples of interest. ### #rm(list=ls()) #<> type <- c("isoform","miRNA","CNV","gene") for(i in seq_along(type)){ cat(type[i],"\n") load(file.path(path,"data",paste0(type[i],"_all.RData")),verbose=TRUE) # TCGA barcode barcode <- rownames(X) code <- sapply(barcode,function(x) strsplit(x,split="-")) code <- as.data.frame(do.call(what=rbind,args=code)) colnames(code) <- c("project","TSS","participant","sample_vial", "portion_analyte","plate","center") code$sample <- substr(code$sample_vial,start=1,stop=2) code$vial <- substr(code$sample_vial,start=3,stop=3) code$portion <- substr(code$portion_analyte,start=1,stop=2) code$analyte <- substr(code$portion_analyte,start=3,stop=3) code$sample_vial <- code$portion_analyte <- NULL # solid tumour (except blood for LAML) solid <- (code$sample=="01" | (y=="TCGA-LAML" & code$sample=="03")) X <- X[solid,] y <- y[solid] # unique samples unique <- !duplicated(substr(rownames(X),start=1,stop=12)) X <- X[unique,] y <- y[unique] save(list=c("y","X"),file=file.path(path,"data",paste0(type[i],"_sub.RData"))) } # isoform: n=9'794, p=194'595, k=32 # miRNA: n=9'794, p=1'881, k=32 # gene: n=9'830, p=19'602, k=33 # CNV: n=10'578, p=10'000, k=33 ## Understanding barcodes: # overview: https://wiki.nci.nih.gov/display/TCGA/TCGA+barcode # details: https://gdc.cancer.gov/resources-tcga-users/tcga-code-tables ## Understanding replicate samples: # http://gdac.broadinstitute.org/runs/sampleReports/latest/READ_Replicate_Samples.html ``` # Analysis Last data analysis started on 2019-04-12 (R version 3.5.3). ```{r do_predict,eval=FALSE} ### Analysing the TCGA data. ### #rm(list=ls()) #<> for(type in c("miRNA","isoform","CNV","gene")){ library(Matrix) for(k in c(207,sample(528))){ set.seed(k); cat(k," ") if(type %in% c("isoform","miRNA") & k > 496){next} if(type %in% c("CNV","gene") & k > 528){next} # searching for missing cancer-cancer combinations rm(list=setdiff(ls(),c("k","type","path","save","file.exists","file.remove"))); gc() file0 <- paste0("results/",type,"_start_",k,".RData") file1 <- paste0("results/",type,"_loss_",k,".RData") if(file.exists(file0)||file.exists(file1)){next} save(object=k,file=file0) load(paste0("data/",type,"_sub.RData")) # indicating the cancer-cancer combination cancer <- substring(text=unique(y),first=6) comb <- utils::combn(x=cancer,m=2) select <- paste0("TCGA-",comb[,k]) y <- ifelse(y==select[1],1,ifelse(y==select[2],0,NA)) rm(cancer,select) # removing other cancer types cond <- !is.na(y) y <- y[cond] X <- X[cond,] rm(cond) # pre-processing if(type %in% c("isoform","miRNA")){ x <- palasso:::.prepare(X,cutoff="zero") } else if(type=="gene"){ x <- palasso:::.prepare(X,cutoff="knee") } else if(type=="CNV"){ x <- list(X=X,Z=sign(X)) x <- lapply(x,function(x) scale(x)) attributes(x)$info <- data.frame(n=nrow(X),p=ncol(X),prop=mean(x$Z==0)) } rm(X) # cross-validation loss <- tryCatch(expr=palasso:::.predict(y=y,X=x,nfolds.int=10),error=function(e) palasso:::.predict(y=y,X=x,nfolds.int=10)) # information loss$info <- cbind(k=k, y0=comb[2,k], y1=comb[1,k], n0=sum(y==0), n1=sum(y==1), attributes(x)$info, loss$info) # refit object <- palasso::palasso(y=y,X=x,nfolds=10,family="binomial",standard=TRUE,elastic=TRUE,shrink=TRUE) model <- c(names(object),"elastic", paste0("paired.",c("adaptive","standard","combined"))) for(max in c(10,50,Inf)){ temp <- list() temp$nzero <- data.frame(model=model,x=NA,z=NA) for(i in seq_along(model)){ coef <- palasso:::coef.palasso(object=object,model=model[i],max=max) temp$nzero$x[i] <- sum(coef$x!=0) temp$nzero$z[i] <- sum(coef$z!=0) } temp$select <- palasso:::subset.palasso(x=object,max=max, model="paired.adaptive")$palasso$select temp$weights <- palasso:::weights.palasso(object=object,max=max, model="paired.adaptive") temp$coef <- palasso:::coef.palasso(object=object,max=max, model="paired.adaptive") loss[[paste0("fit",max)]] <- temp } save(object=loss,file=file1) file.remove(file0) } index <- sum(grepl(dir(),pattern="sessionInfo")) sink(paste0("sessionInfo",index+1,".txt")) date() utils::sessionInfo() devtools::session_info() sink() } ``` ```{r collect} # The function <> loads all files from PATH including PATTERN in the file name, loads OBJECT into a list, and executes a function call. #<> # path: folder # pattern: character, or NULL (all files) # object: character vector, or NULL (all objects) # what: function call collect <- function(path=getwd(),pattern="",object=NULL,what="rbind"){ OBJECT = object # identify files files <- dir(path) files <- files[grepl(x=files,pattern=pattern)] files <- files[grepl(x=files,pattern=".RData")] number <- gsub(pattern=paste0(pattern,"|.RData"),replacement="",x=files) files <- files[order(as.numeric(number))] # trial1 names <- gsub(pattern=".RData",replacement="",x=files) # trial1 if(length(files)==0){stop("Invalid datasets.")} # load data all <- list() for(i in seq_along(files)){ x <- load(file.path(path,files[i])) x <- eval(parse(text=x)) if(is.null(OBJECT)){ all[[i]] <- x names(all)[i] <- names[i] } else { for(j in seq_along(OBJECT)){ all[[OBJECT[j]]][[i]] <- x[[OBJECT[j]]] names(all[[OBJECT[j]]])[i] <- names[i] } } } # fuse data if(is.null(OBJECT)){ all <- do.call(what=what,args=all) } else { all <- lapply(all,function(x) do.call(what=what,args=x)) } return(all) } LOSS <- list() type <- c("gene","isoform","miRNA","CNV") #type <- "miRNA" for(i in seq_along(type)){ LOSS[[type[i]]] <- collect(path="results", pattern=paste0(type[i],"_loss_"), object=c("deviance","auc","class","info", paste0("fit",c(10,50,Inf)))) } for(i in seq_along(LOSS)){ for(j in 1:3){ colnames(LOSS[[i]][[j]])[colnames(LOSS[[i]][[j]])=="paired.adaptive"] <- "paired" } } #type <- "gene" #a <- LOSS[[type]]$deviance[rownames(LOSS[[type]]$deviance)=="10","paired"] #b <- LOSS[[type]]$deviance[rownames(LOSS[[type]]$deviance)=="10","elastic"] #mean(a> #<> row <- c("gene","isoform","miRNA","CNV") col <- c("10","Inf") lay <- c("standard_x","standard_z","standard_xz", "adaptive_x","adaptive_z","adaptive_xz", "elastic") # added "elastic" M <- array(NA,dim=c(length(row),length(col),length(lay)),dimnames=list(row,col,lay)) for(i in seq_along(row)){ loss <- LOSS[[row[i]]][c("info","deviance")] y0 <- as.character(loss$info$y0) y1 <- as.character(loss$info$y1) cancer <- sort(unique(c(y0,y1))) Z <- palasso:::.design(x=cancer) for(j in seq_along(col)){ # differences cond <- rownames(loss$deviance)==col[j] for(k in seq_along(lay)){ fill <- loss$deviance[cond,lay[k]] - loss$deviance[cond,"paired"] X <- matrix(NA,nrow=length(cancer),ncol=length(cancer), dimnames=list(cancer,cancer)) X[cbind(y0,y1)] <- X[cbind(y1,y0)] <- fill X[lower.tri(X)] <- NA # p-values pvalue <- rep(NA,times=max(Z)) for(l in seq_len(max(Z))){ x <- as.numeric(X[Z==l]) if(col[j]=="10"){ alternative <- "greater" # Never use "two.sided"! } if(col[j]=="Inf"){ alternative <- "less" # Never use "two.sided"! } pvalue[l] <- stats::wilcox.test(x=x,alternative=alternative, exact=FALSE)$p.value } # Simes M[i,j,k] <- palasso:::.combine(pvalue,method="simes") } } } # Table SIG: significance constraint <- "10" table <- format(M[,constraint,1:6],digits=1,scientific=FALSE) for(i in seq_len(nrow(table))){ for(j in seq_len(ncol(table))){ if(M[i,constraint,j]>=0.05){ table[i,j] <- paste0("{\\color{gray}{",table[i,j],"}}") } } } one <- c("","\\text{standard}","","","\\text{adaptive}","") two <- paste0("\\text{",c("x","z","xz","x","z","xz"),"}") rownames(table) <- paste0("\\text{",rownames(table),"}") table <- rbind(one,two,table,deparse.level=0) rownames(table)[1] <- "~" xtable <- xtable::xtable(table,align=c("r","|","c","c","c","|","c","c","c")) xtable::print.xtable(xtable,type="latex",include.colnames=FALSE,sanitize.text.function=identity) ``` ```{r do_elastic,eval=eval} ### Comparison with elastic net. ### #rm(list=ls()) #<> #<> row <- c("gene","isoform","miRNA","CNV") col <- c("10","50","Inf") better <- worse <- less <- matrix(NA,nrow=length(row),ncol=length(col), dimnames=list(row,col)) for(i in seq_along(row)){ for(j in seq_along(col)){ # proportion of improvements (cross-validation) cond <- rownames(LOSS[[row[i]]]$deviance)==col[j] loss <- LOSS[[row[i]]]$deviance[cond,c("paired","elastic")] better[i,j] <- round(mean(loss[,"paired"]loss[,"elastic"]),digits=2) # average difference in nzero (refitted models) df_paired <- apply(LOSS[[row[i]]][[paste0("fit",col[j])]],1,function(x) sum(x$nzero[x$nzero[,"model"]=="paired.adaptive",c("x","z")])) df_elastic <- apply(LOSS[[row[i]]][[paste0("fit",col[j])]],1,function(x) sum(x$nzero[x$nzero[,"model"]=="elastic95",c("x","z")])) df_diff <- df_elastic-df_paired less[i,j] <- round(mean(df_diff),digits=2) #graphics::hist(df_diff,main=paste(row[i],col[j]),xlim=c(-1,1)*max(abs(df_diff))) } } better worse less ``` ```{r do_refit,eval=eval} ### Analysing the refitted models. ### #rm(list=ls()) #<> #<> # Table SEL: selected model nzero <- paste0("fit",c(5,10,Inf)) model <- c(paste0("standard_",c("x","z","xz")), paste0("adaptive_",c("x","z","xz")), "between_xz","within_xz") type <- c("gene","isoform","miRNA","CNV") table <- array(NA,dim=c(length(nzero),length(model),length(type)), dimnames=list(nzero,model,type)) for(i in seq_along(nzero)){ for(j in seq_along(model)){ for(k in seq_along(type)){ sub <- LOSS[[type[k]]][[nzero[i]]] table[i,j,k] <- sum(sub[,"select"]==model[j]) } } } colSums(table["fit10",,]) # CHECK WHETHER COMPLETE! table <- round(prop.table(table["fit10",,],margin=2),digits=2) table <- t(table) table <- table[,apply(table,2,function(x) any(x!=0))] rownames(table) <- paste0("\\text{",rownames(table),"}") xtable <- xtable::xtable(table,align=c("r","|","c","c","c","c")) xtable::print.xtable(xtable,type="latex",include.colnames=FALSE,sanitize.text.function=identity) # selected weights and covariates type <- c("gene","isoform","miRNA","CNV") group <- c("x","z") model <- c(paste0("standard_",c("x","z","xz")), paste0("adaptive_",c("x","z","xz")), "paired.adaptive","elastic") # added "elastic" , paste0("elastic",c(100,75,50,25)) weights10 <- weightsInf <- matrix(NA,nrow=length(group),ncol=length(type), dimnames=list(group,type)) coef10 <- coefInf <- array(NA,dim=c(length(group),length(type),length(model)), dimnames=list(group,type,model)) for(i in seq_along(group)){ for(j in seq_along(type)){ weights10[,j] <- rowMeans(sapply(LOSS[[type[j]]]$fit10[,"weights"],colMeans)) weightsInf[,j] <- rowMeans(sapply(LOSS[[type[j]]]$fitInf[,"weights"],colMeans)) for(k in seq_along(model)){ coef10[i,j,k] <- mean(sapply(LOSS[[type[j]]]$fit10[,"nzero"], function(x) sum(x[x$model==model[k],group[i]]))) coefInf[i,j,k] <- mean(sapply(LOSS[[type[j]]]$fitInf[,"nzero"], function(x) sum(x[x$model==model[k],group[i]]))) } } } # coef10["x",,]+coef10["z",,] # with sparsity constraint round(prop.table(weights10,margin=2),2) round(prop.table(coef10[,,"paired.adaptive"],margin=2),2) round(colSums(coef10[,,"paired.adaptive"]),2) # natural sparsity round(prop.table(weightsInf,margin=2),2) round(prop.table(coefInf[,,"paired.adaptive"],margin=2),2) round(colSums(coefInf[,,"paired.adaptive"]),2) round(colSums(coefInf[,,"elastic"])/colSums(coefInf[,,"paired.adaptive"]),1) # multiple nzero of elastic net and paired lasso # Table NSC: number of non-zero coefficients table <- round(coefInf["x",,]+coefInf["z",,]) colnames(table)[7] <- "paired" one <- c("","\\text{standard}","","","\\text{adaptive}","","\\text{paired}","\\text{elastic}") two <- paste0("\\text{",c("x","z","xz","x","z","xz","xz","xz"),"}") rownames(table) <- paste0("\\text{",rownames(table),"}") table <- rbind(one,two,table,deparse.level=0) rownames(table)[1] <- "~" xtable <- xtable::xtable(table,align=c("r","|","c","c","c","|","c","c","c","|","c","|","c")) xtable::print.xtable(xtable,type="latex",include.colnames=FALSE,sanitize.text.function=identity) ``` # Figures ```{r figure_CSW,fig.height=2,fig.cap="__Figure CSW:__ Weighting schemes. Each covariate pair ($y$-axis) receives weights for both parts ($x$-axis), here for simulated data."} ### FIGURE CSW ### #rm(list=ls()) #<> set.seed(1) overfit <- TRUE # simulate n <- 10 cx <- stats::rbeta(n=n,shape1=0.9,shape2=1) cz <- stats::rbeta(n=n,shape1=0.4,shape2=0.9) # collection x <- list() y <- list() # adaptive weights (X only) x[[1]] <- rep(1,times=n) if(overfit){x[[1]] <- cx} y[[1]] <- rep(0,times=n) # adaptive weights (Z only) x[[2]] <- rep(0,times=n) y[[2]] <- rep(1,times=n) if(overfit){y[[2]] <- cz} # adaptive weights (X and Z) x[[3]] <- y[[3]] <- rep(0.5,times=n) if(overfit){x[[3]] <- cx} if(overfit){y[[3]] <- cz} # within-pair weights x[[4]] <- cx^2/(cx+cz) y[[4]] <- cz^2/(cx+cz) # visualisation graphics::par(mfrow=c(1,4),mar=c(4.5,0.5,0.5,0.5),oma=c(0,2,0,0)) for(i in seq_len(4)){ palasso:::plot_pairs(x=x[[i]],y=y[[i]],lwd=4) if(i==1){ graphics::mtext(text="covariate pair",side=2,line=1) } } ``` ```{r figure_DIA,fig.height=3,fig.cap="__Figure DIA__: Sample size flowchart. \\textsc{tcga} provides suitable isomi\\textsc{r} data for $9\\,794$ samples (left), from $32$ cancer types (centre), forming $496$ cancer-cancer combinations (right). Each sample appears in $31$ combinations."} ### FIGURE DIA ### #rm(list=ls()) #<> ellipse <- function(x,y,a=0.2,b=0.25,border=NA){ n <- max(c(length(x),length(y))) if(length(x)==1){x <- rep(x,times=n)} if(length(y)==1){y <- rep(y,times=n)} if(length(border)==1){border <- rep(border,times=n)} for(i in seq_len(n)){ angle <- seq(from=0,to=2*pi,length=100) xs <- x[i] + a * cos(angle) ys <- y[i] + b * sin(angle) graphics::polygon(x=xs,y=ys,col=grey(0.9),border=border[i]) } } cancer <- c("ACC","BLCA","BRCA","UVM") number <- c(80,409,1078,80) col <- grDevices::rgb(red=0,green=0,blue=sample(seq(from=75,to=255,length.out=length(number))),maxColorValue=255) lwd <- log(2*number)-2 lwd = pmax(5*number/max(number),1) x1 <- 1.3; x2 <- 2; x3 <- 3 set.seed(1) # first layer (bubble) graphics::plot.new() graphics::par(mar=c(0,0,0,0)) graphics::plot.window(xlim=c(1.1,3.3),ylim=c(-1.6,1.6)) ellipse(x=x1,y=0) graphics::text(x=x1,y=0,labels="TCGA",font=2,adj=c(0.5,0)) graphics::text(x=x1,y=0,labels="n=9794",adj=c(0.5,1.2),cex=0.9) # second layer (colon) y1 <- seq(from=1,to=-1,length.out=length(cancer)+1) graphics::text(x=x2,y=y1[length(cancer)],labels="...",font=2,srt=90) y1 <- y1[-length(cancer)] # first-second layer (connect) graphics::segments(x0=x1+0.2,y0=0,x1=x2-0.2,y1=y1,lwd=lwd,col=col) # second layer (bubble) ellipse(x=x2,y=y1,a=0.2,b=0.2) graphics::text(x=x2,y=y1,labels=cancer,font=2,col=col,adj=c(0.5,0)) graphics::text(x=x2,y=y1,labels=paste0("n=",number,""),adj=c(0.5,1.2),cex=0.9) comb <- utils::combn(x=seq_along(y1),m=2) # third layer (colon) y2 <- seq(from=1.5,to=-1.5,length.out=ncol(comb)+1) graphics::text(x=x3,y=y2[ncol(comb)],labels="...",font=2,srt=90) y2 <- y2[-ncol(comb)] # second-third layer (connect) graphics::segments(x0=2.2,y0=y1[comb[1,]],x1=2.7,y1=y2,lwd=lwd[comb[1,]],col=col[comb[1,]]) graphics::segments(x0=2.2,y0=y1[comb[2,]],x1=2.7,y1=y2,lwd=lwd[comb[2,]],col=col[comb[2,]]) # third layer (bubble) ellipse(x=x3,y=y2,a=0.3,b=0.22) graphics::text(x=x3,y=y2,labels=paste0(cancer[comb[1,]]," "), font=2,col=col[comb[1,]],adj=c(1,0)) graphics::text(x=x3,y=y2,labels=paste0(" ",cancer[comb[2,]]), font=2,col=col[comb[2,]],adj=c(0,0)) graphics::text(x=x3,y=y2,labels=":",font=2,adj=c(0.5,0)) labels <- apply(comb,2,function(x) sum(number[x])) labels <- paste0("n=",labels,"") graphics::text(x=x3,y=y2,labels=labels,adj=c(0.5,1.2),cex=0.9) ``` ```{r figure_CLA,fig.height=5,fig.width=6,message=FALSE,fig.cap="__Figure CLA:__ Predictive performance for genes, isomi\\textsc{r}s, mi\\textsc{rna}s and \\textsc{cnv}s (from top to bottom). The bar charts (left) count how often the paired lasso leads to a lower (dark) or higher (bright) deviance than the competing model. The box plots (right) show how much lower (dark) or higher (bright) the deviance is."} ### FIGURE CLA ### #rm(list=ls()) #<> graphics::par(mfrow=c(4,2),oma=c(0,0,0,0),mar=c(2.1,3.5,0.5,0.5)) for(type in c("gene","isoform","miRNA","CNV")){ loss <- LOSS[[type]][c("deviance","auc","class")] choice <- "paired" loss <- lapply(loss,function(x) x[,c(paste0("standard_",c("x","z","xz")),paste0("adaptive_",c("x","z","xz")),choice)]) for(constraint in c("10")){ # c("5","10","Inf") # change sub <- lapply(loss,function(x) x[rownames(x)==constraint,]) palasso:::plot_score(sub$deviance,choice=choice) change <- sub$deviance[,7]-sub$deviance[,-7] palasso:::plot_box(change,ylab="change",zero=TRUE,choice=NA) # info info <- list() info$select <- names(which.min(apply(sub$deviance,2,median)[-7])) info$DEV_paired <- median(sub$deviance[,choice]) info$DEV_select <- median(sub$deviance[,info$select]) info$improve <- mean(sub$deviance[,info$select]>sub$deviance[,choice]) info$AUC_paired <- median(sub$auc[,choice]) info$CLASS_paired <- median(sub$class[,choice]) print(as.data.frame(info)) # important } } ``` ```{r figure_DEC,fig.height=4,fig.width=6,fig.cap="__Figure DEC:__ Model convergence for genes (top left), isomi\\textsc{r}s (top right), mi\\textsc{rna}s (bottom left) and \\textsc{cnv}s (bottom right). The median deviances ($y$-axis) of the standard (dotted), adaptive (dashed) and paired (solid) lasso converge as the sparsity constraint ($x$-axis) increases."} ### FIGURE DEC ### #rm(list=ls()) #<> #graphics::par(oma=c(1.0,1.0,0,0),mar=c(1.5,3.0,0.5,0.5),mfrow=c(1,1)) graphics::par(oma=c(1.0,1.0,0,0),mar=c(1.5,3.0,0.5,0.5),mfrow=c(2,2)) for(type in c("gene","isoform","miRNA","CNV")){ models <- c(paste0("standard_",c("x","z","xz")), paste0("adaptive_",c("x","z","xz")),"paired") constraint <- c("3","4","5","10","15","20","25","50","Inf") loss <- LOSS[[type]]["deviance"] loss <- lapply(loss,function(x) x[,models]) table <- matrix(NA,nrow=length(constraint),ncol=length(models), dimnames=list(constraint,models)) for(i in seq_along(constraint)){ sub <- lapply(loss,function(x) x[rownames(x)==constraint[i],]) table[i,] <- apply(sub$deviance,2,median) } # table <- log(table) graphics::plot.new() graphics::plot.window(xlim=c(1,length(constraint)),ylim=range(table)) graphics::box() constraint[constraint=="Inf"] <- "n" graphics::axis(side=2) graphics::axis(side=1,at=seq_along(constraint),labels=constraint,tick=FALSE,line=-1) for(k in c(1,2)){ for(i in seq_along(models)){ lty <- ifelse(i%in%c(1,2,3),3,ifelse(i%in%c(4,5,6),2,1)) col <- ifelse(i==7,"#00007F","#FF3535") pch <- ifelse(i%in%c(1,4),"x",ifelse(i%in%c(2,5),"z",1)) if(k==1){ graphics::lines(table[,i],col=col,lty=lty,lwd=2) graphics::points(table[,i],col="white",pch=16,cex=1.2) } else { graphics::points(table[,i],col=col,pch=1,font=2) } } } } graphics::title(ylab="deviance",line=0.0,outer=TRUE) graphics::title(xlab="sparsity constraint",ylab="deviance",line=0.0,outer=TRUE) ``` ```{r figure_CNV,fig.height=1.25,fig.width=6,fig.cap="__Figure CNV:__ Predictive performance for \\textsc{cnv}s. The box plots show how much the paired lasso improves (dark) or deteriorates (bright) the \\textsc{auc} (left) and misclassification rate (right) of the competing models."} ### FIGURE CNV ### #rm(list=ls()) #<> graphics::par(oma=c(0,0,0,0),mar=c(2.1,3.5,0.5,0.5)) graphics::layout(matrix(c(1,1,2,2),nrow=1)) loss <- LOSS[[type]][c("deviance","auc","class")] loss <- lapply(loss,function(x) x[rownames(x)=="10",]) model <- c(paste0("standard_",c("x","z","xz")), paste0("adaptive_",c("x","z","xz"))) diff <- loss$auc[,"paired"]-loss$auc[,model] palasso:::plot_box(diff,zero=TRUE,invert=TRUE,ylab="change") diff <- loss$class[,"paired"]-loss$class[,model] palasso:::plot_box(diff,zero=TRUE,ylab="change") ``` ```{r figure_MAP,fig.height=4,fig.width=4,fig.cap="__Figure MAP:__ Cross-validated \\textsc{auc} for \\textsc{cnv}s. Each cell represents one cancer-cancer combination (row, column). The colour indicates whether the paired lasso leads to a low (dark) or high (bright) \\textsc{auc}."} ### FIGURE MAP ### #rm(list=ls()) #<> loss <- LOSS[["CNV"]][c("info","auc")] cancer <- sort(unique(c(levels(loss$info$y0),levels(loss$info$y1)))) X <- matrix(NA,nrow=length(cancer),ncol=length(cancer),dimnames=list(cancer,cancer)) #Z <- palasso:::.design(x=cancer) y0 <- as.character(loss$info$y0) y1 <- as.character(loss$info$y1) X[cbind(y0,y1)] <- X[cbind(y1,y0)] <- loss$auc[rownames(loss$auc)=="10","paired"] graphics::par(mar=c(0.5,3.0,3.0,0.5)) dimnames(X) <- lapply(dimnames(X),function(x) paste0(" ",x," ")) palasso:::plot_table(X=X,margin=-1,labels=FALSE,las=2,cex=0.7) #sort(rowMeans(X,na.rm=TRUE),decreasing=TRUE)[1:2] # keep! ``` ```{r figure_COM,fig.height=3.5,fig.width=4,fig.cap="__Figure COM:__ Group assignment for isomi\\textsc{r}s. Given $32$ cancer types, this matrix shows the assignment of $496$ dependent pairs to $31$ groups of $16$ independent pairs, with each symbol representing one group."} ### FIGURE COM ### # 32 cancer types for isoform and miRNA # 33 cancer types for gene and CNV #rm(list=ls()) #<> for(type in c("miRNA")){ cancer <- sort(unique(as.character(unlist(LOSS[[type]]$info[,c("y0","y1")])))) n <- length(cancer) z <- as.numeric(palasso:::.design(x=n)) x <- rep(seq_len(n),each=n) y <- rep(seq(from=n,to=1,by=-1),times=n) pch <- z pch[pch==0] <- NA pex <- c(".","O","*","+","o","-","'","x") # colour base <- grDevices::colorRampPalette(colors=c('darkblue','blue','red','darkred'))(n) col <- rep(NA,times=length(z)) col[z==0] <- "white" for(i in seq_len(n)){ col[z==i] <- base[i] } graphics::par(mfrow=c(1,1),mar=c(0,0,2,2)) graphics::plot.new() graphics::plot.window(xlim=c(1,n),ylim=c(1,n)) graphics::points(x=x[pch<=25],y=y[pch<=25], pch=pch[pch<=25],col=col[pch<=25],cex=0.9) graphics::points(x=x[pch>25],y=y[pch>25], pch=pex[(pch-25)[pch>25]],col=col[pch>25],cex=0.9) graphics::segments(x0=0,x1=n+1,y0=n+1) graphics::segments(x0=n+1,y0=n+1,y1=0) graphics::segments(x0=0,x1=n+1,y0=n+1,y1=0,lty=2) graphics::mtext(text=cancer,side=3,at=1:n,las=2,cex=0.7) graphics::mtext(text=cancer,side=4,at=n:1,las=2,cex=0.7) } ``` *** *** ***