Mercurial > repos > modencode-dcc > idr_package
view functions-all-clayton-12-13.r @ 10:e9eb89c2e71d draft
Uploaded
author | modencode-dcc |
---|---|
date | Thu, 17 Jan 2013 15:59:14 -0500 |
parents | 5e6efd5f3567 |
children |
line wrap: on
line source
# revised on 2-20-10 # - fix error in pass.structure: reverse rank.combined, so that big sig.value # are ranked with small numbers (1, 2, ...) # - fix error on get.ez.tt.all: get ez.cutoff from sorted e.z # # modified EM procedure to compute empirical CDF more precisely - 09/2009 # this file contains the functions for # 1. computing the correspondence profile (upper rank intersection and derivatives) # 2. inference of copula mixture model # # It also has functions for # 1. reading peak caller results # 2. processing and matching called peaks # 3. plotting results ################ read peak caller results # process narrow peak format # some peak callers may not report q-values, p-values or fold of enrichment # need further process before comparison # # stop.exclusive: Is the basepair of peak.list$stop exclusive? In narrowpeak and broadpeak format they are exclusive. # If it is exclusive, we need subtract peak.list$stop by 1 to avoid the same basepair being both a start and a stop of two # adjacent peaks, which creates trouble for finding correct intersect process.narrowpeak <- function(narrow.file, chr.size, half.width=NULL, summit="offset", stop.exclusive=T, broadpeak=F){ aa <- read.table(narrow.file) if(broadpeak){ bb.ori <- data.frame(chr=aa$V1, start=aa$V2, stop=aa$V3, signal.value=aa$V7, p.value=aa$V8, q.value=aa$V9) }else{ bb.ori <- data.frame(chr=aa$V1, start=aa$V2, stop=aa$V3, signal.value=aa$V7, p.value=aa$V8, q.value=aa$V9, summit=aa$V10) } if(summit=="summit"){ bb.ori$summit <- bb.ori$summit-bb.ori$start # change summit to offset to avoid error when concatenating chromosomes } bb <- concatenate.chr(bb.ori, chr.size) #bb <- bb.ori # remove the peaks that has the same start and stop value bb <- bb[bb$start != bb$stop,] if(stop.exclusive==T){ bb$stop <- bb$stop-1 } if(!is.null(half.width)){ bb$start.ori <- bb$start bb$stop.ori <- bb$stop # if peak is narrower than the specified window, stay with its width # otherwise chop wider peaks to specified width width <- bb$stop-bb$start +1 is.wider <- width > 2*half.width if(summit=="offset" | summit=="summit"){ # if summit is offset from start bb$start[is.wider] <- bb$start.ori[is.wider] + bb$summit[is.wider]-half.width bb$stop[is.wider] <- bb$start.ori[is.wider] + bb$summit[is.wider]+half.width } else { if(summit=="unknown"){ bb$start[is.wider] <- bb$start.ori[is.wider]+round(width[is.wider]/2) - half.width bb$stop[is.wider] <- bb$start.ori[is.wider]+round(width[is.wider]/2) + half.width } } } bb <- clean.data(bb) invisible(list(data.ori=bb.ori, data.cleaned=bb)) } # clean data # and concatenate chromosomes if needed clean.data <- function(adata){ # remove the peaks that has the same start and stop value adata <- adata[adata$start != adata$stop,] # if some stops and starts are the same, need fix them stop.in.start <- is.element(adata$stop, adata$start) n.fix <- sum(stop.in.start) if(n.fix >0){ print(paste("Fix", n.fix, "stops\n")) adata$stop[stop.in.start] <- adata$stop[stop.in.start]-1 } return(adata) } # concatenate peaks # peaks: the dataframe to have all the peaks # chr.file: the file to keep the length of each chromosome # chr files should come from the species that the data is from #concatenate.chr <- function(peaks, chr.size){ # chr.size <- read.table(chr.file) # chr.o <- order(chr.size[,1]) # chr.size <- chr.size[chr.o,] # # chr.shift <- cumsum(c(0, chr.size[-nrow(chr.size),2])) # chr.size.cum <- data.frame(chr=chr.size[,1], shift=chr.shift) # # for(i in 1:nrow(chr.size)){ # is.in <- as.character(peaks$chr) == as.character(chr.size.cum$chr[i]) # if(sum(is.in)>0){ # peaks[is.in,]$start <- peaks[is.in,]$start + chr.size.cum$shift[i] # peaks[is.in,]$stop <- peaks[is.in,]$stop + chr.size.cum$shift[i] # } # } # # invisible(peaks) #} # concatenate peaks # peaks: the dataframe to have all the peaks # chr.file: the file to keep the length of each chromosome # chr files should come from the species that the data is from concatenate.chr <- function(peaks, chr.size){ # chr.size <- read.table(chr.file) chr.o <- order(chr.size[,1]) chr.size <- chr.size[chr.o,] chr.shift <- cumsum(c(0, chr.size[-nrow(chr.size),2])) chr.size.cum <- data.frame(chr=chr.size[,1], shift=chr.shift) peaks$start.ori <- peaks$start peaks$stop.ori <- peaks$stop for(i in 1:nrow(chr.size)){ is.in <- as.character(peaks$chr) == as.character(chr.size.cum$chr[i]) if(sum(is.in)>0){ peaks[is.in,]$start <- peaks[is.in,]$start + chr.size.cum$shift[i] peaks[is.in,]$stop <- peaks[is.in,]$stop + chr.size.cum$shift[i] } } invisible(peaks) } deconcatenate.chr <- function(peaks, chr.size){ chr.o <- order(chr.size[,1]) chr.size <- chr.size[chr.o,] chr.shift <- cumsum(c(0, chr.size[-nrow(chr.size),2])) chr.size.cum <- data.frame(chr=chr.size[,1], shift=chr.shift) peaks$chr <- rep(NA, nrow(peaks)) for(i in 1:(nrow(chr.size.cum)-1)){ is.in <- peaks$start > chr.size.cum[i,2] & peaks$start <= chr.size.cum[i+1, 2] if(sum(is.in)>0){ peaks[is.in,]$start <- peaks[is.in,]$start - chr.size.cum[i,2] peaks[is.in,]$stop <- peaks[is.in,]$stop - chr.size.cum[i,2]+1 peaks[is.in,]$chr <- chr.size[i,1] } } if(i == nrow(chr.size.cum)){ is.in <- peaks$start > chr.size.cum[i, 2] if(sum(is.in)>0){ peaks[is.in,]$start <- peaks[is.in,]$start - chr.size.cum[i,2] peaks[is.in,]$stop <- peaks[is.in,]$stop - chr.size.cum[i,2]+1 peaks[is.in,]$chr <- chr.size[i,1] } } invisible(peaks) } ################ preprocessing peak calling output # # read two calling results and sort by peak starting locations, # then find overlap between peaks # INPUT: # rep1: the 1st replicate # rep2: the 2nd replicate # OUTPUT: # id1, id2: the labels for the identified peaks on the replicates find.overlap <- function(rep1, rep2){ o1 <- order(rep1$start) rep1 <- rep1[o1,] o2 <- order(rep2$start) rep2 <- rep2[o2,] n1 <- length(o1) n2 <- length(o2) # assign common ID to peaks id1 <- rep(0, n1) # ID assigned on rep1 id2 <- rep(0, n2) # ID assigned on rep2 id <- 1 # keep track common id's # check if two replicates overlap with each other i <- 1 j <- 1 while(i <= n1|| j <= n2){ # && (id1[n1] ==0 || id2[n2] ==0) # if one list runs out if(i > n1 && j < n2){ j <- j+1 id2[j] <- id id <- id +1 next } else{ if(j > n2 && i < n1){ i <- i+1 id1[i] <- id id <- id +1 next } else { if(i >= n1 && j >=n2) break } } # if not overlap if(!(rep1$start[i] <= rep2$stop[j] && rep2$start[j] <= rep1$stop[i])){ # at the start of loop, when both are not assigned an ID # the one locates in front is assigned first if(id1[i] ==0 && id2[j]==0){ if(rep1$stop[i] < rep2$stop[j]){ id1[i] <- id } else { id2[j] <- id } } else { # in the middle of the loop, when one is already assigned # The one that has not assigned gets assigned # if(id1[i] ==0){ # id1[i] is not assigned # id1[i] <- id # } else { # id2[i] is not assigned # id2[j] <- id # } # order the id according to location if(rep1$stop[i] <= rep2$stop[j]){ id1[i] <- max(id2[j], id1[i]) id2[j] <- id } else { if(rep1$stop[i] > rep2$stop[j]){ id2[j] <- max(id1[i], id2[j]) id1[i] <- id } } } id <- id +1 } else { # if overlap if(id1[i] == 0 && id2[j] == 0){ # not assign label yet id1[i] <- id id2[j] <- id id <- id +1 } else { # one peak is already assigned label, the other is 0 id1[i] <- max(id1[i], id2[j]) # this is a way to copy the label of the assigned peak without knowing which one is already assigned id2[j] <- id1[i] # syncronize the labels } } if(rep1$stop[i] < rep2$stop[j]){ i <- i+1 } else { j <- j+1 } } invisible(list(id1=id1, id2=id2)) } # Impute the missing significant value for the peaks called only on one replicate. # value # INPUT: # rep1, rep2: the two peak calling output # id1, id2: the IDs assigned by function find.overlap, vectors # If id1[i]==id2[j], peak i on rep1 overlaps with peak j on rep2 # p.value.impute: the significant value to impute for the missing peaks # OUTPUT: # rep1, rep2: peaks ordered by the start locations with imputed peaks # id1, id2: the IDs with imputed peaks fill.missing.peaks <- function(rep1, rep2, id1, id2, p.value.impute){ # rep1 <- data.frame(chr=rep1$chr, start=rep1$start, stop=rep1$stop, sig.value=rep1$sig.value) # rep2 <- data.frame(chr=rep2$chr, start=rep2$start, stop=rep2$stop, sig.value=rep2$sig.value) o1 <- order(rep1$start) rep1 <- rep1[o1,] o2 <- order(rep2$start) rep2 <- rep2[o2,] entry.in1.not2 <- !is.element(id1, id2) entry.in2.not1 <- !is.element(id2, id1) if(sum(entry.in1.not2) > 0){ temp1 <- rep1[entry.in1.not2, ] # impute sig.value temp1$sig.value <- p.value.impute temp1$signal.value <- p.value.impute temp1$p.value <- p.value.impute temp1$q.value <- p.value.impute rep2.filled <- rbind(rep2, temp1) id2.filled <- c(id2, id1[entry.in1.not2]) } else { id2.filled <- id2 rep2.filled <- rep2 } if(sum(entry.in2.not1) > 0){ temp2 <- rep2[entry.in2.not1, ] # fill in p.values to 1 temp2$sig.value <- p.value.impute temp2$signal.value <- p.value.impute temp2$p.value <- p.value.impute temp2$q.value <- p.value.impute # append to the end rep1.filled <- rbind(rep1, temp2) id1.filled <- c(id1, id2[entry.in2.not1]) } else { id1.filled <- id1 rep1.filled <- rep1 } # sort rep1 and rep2 by the same id o1 <- order(id1.filled) rep1.ordered <- rep1.filled[o1, ] o2 <- order(id2.filled) rep2.ordered <- rep2.filled[o2, ] invisible(list(rep1=rep1.ordered, rep2=rep2.ordered, id1=id1.filled[o1], id2=id2.filled[o2])) } # Merge peaks with same ID on the same replicates # (They are generated if two peaks on rep1 map to the same peak on rep2) # need peak.list have 3 columns: start, stop and sig.value merge.peaks <- function(peak.list, id){ i <- 1 j <- 1 dup.index <- c() sig.value <- c() start.new <- c() stop.new <- c() id.new <- c() # original data chr <- c() start.ori <- c() stop.ori <- c() signal.value <- c() p.value <- c() q.value <- c() while(i < length(id)){ if(id[i] == id[i+1]){ dup.index <- c(dup.index, i, i+1) # push on dup.index } else { if(length(dup.index)>0){ # pop from dup.index sig.value[j] <- mean(peak.list$sig.value[unique(dup.index)]) # mean of -log(pvalue) start.new[j] <- peak.list$start[min(dup.index)] stop.new[j] <- peak.list$stop[max(dup.index)] id.new[j] <- id[max(dup.index)] signal.value[j] <- mean(peak.list$signal.value[unique(dup.index)]) # mean of -log(pvalue) p.value[j] <- mean(peak.list$p.value[unique(dup.index)]) # mean of -log(pvalue) q.value[j] <- mean(peak.list$q.value[unique(dup.index)]) # mean of -log(pvalue) chr[j] <- as.character(peak.list$chr[min(dup.index)]) start.ori[j] <- peak.list$start.ori[min(dup.index)] stop.ori[j] <- peak.list$stop.ori[max(dup.index)] dup.index <- c() } else { # nothing to pop sig.value[j] <- peak.list$sig.value[i] start.new[j] <- peak.list$start[i] stop.new[j] <- peak.list$stop[i] id.new[j] <- id[i] signal.value[j] <- peak.list$signal.value[i] p.value[j] <- peak.list$p.value[i] q.value[j] <- peak.list$q.value[i] chr[j] <- as.character(peak.list$chr[i]) start.ori[j] <- peak.list$start.ori[i] stop.ori[j] <- peak.list$stop.ori[i] } j <- j+1 } i <- i+1 } data.new <- data.frame(id=id.new, sig.value=sig.value, start=start.new, stop=stop.new, signal.value=signal.value, p.value=p.value, q.value=q.value, chr=chr, start.ori=start.ori, stop.ori=stop.ori) invisible(data.new) } # a wrap function to fill in missing peaks, merge peaks and impute significant values # out1 and out2 are two peak calling outputs pair.peaks <- function(out1, out2, p.value.impute=0){ aa <- find.overlap(out1, out2) bb <- fill.missing.peaks(out1, out2, aa$id1, aa$id2, p.value.impute=0) cc1 <- merge.peaks(bb$rep1, bb$id1) cc2 <- merge.peaks(bb$rep2, bb$id2) invisible(list(merge1=cc1, merge2=cc2)) } # overlap.ratio is a parameter to define the percentage of overlap # if overlap.ratio =0, 1 basepair overlap is counted as overlap # if overlap.ratio between 0 and 1, it is the minimum proportion of # overlap required to be called as a match # it is computed as the overlap part/min(peak1.length, peak2.length) pair.peaks.filter <- function(out1, out2, p.value.impute=0, overlap.ratio=0){ aa <- find.overlap(out1, out2) bb <- fill.missing.peaks(out1, out2, aa$id1, aa$id2, p.value.impute=0) cc1 <- merge.peaks(bb$rep1, bb$id1) cc2 <- merge.peaks(bb$rep2, bb$id2) frag12 <- cbind(cc1$start, cc1$stop, cc2$start, cc2$stop) frag.ratio <- apply(frag12, 1, overlap.middle) frag.ratio[cc1$sig.value==p.value.impute | cc2$sig.value==p.value.impute] <- 0 cc1$frag.ratio <- frag.ratio cc2$frag.ratio <- frag.ratio merge1 <- cc1[cc1$frag.ratio >= overlap.ratio,] merge2 <- cc2[cc2$frag.ratio >= overlap.ratio,] invisible(list(merge1=merge1, merge2=merge2)) } # x[1], x[2] are the start and end of the first fragment # and x[3] and x[4] are the start and end of the 2nd fragment # If there are two fragments, we can find the overlap by ordering the # start and stop of all the ends and find the difference between the middle two overlap.middle <- function(x){ x.o <- x[order(x)] f1 <- x[2]-x[1] f2 <- x[4]-x[3] f.overlap <- abs(x.o[3]-x.o[2]) f.overlap.ratio <- f.overlap/min(f1, f2) return(f.overlap.ratio) } ####### ####### compute correspondence profile ####### # compute upper rank intersection for one t # tv: the upper percentile # x is sorted by the order of paired variable comp.uri <- function(tv, x){ n <- length(x) qt <- quantile(x, prob=1-tv[1]) # tv[1] is t # sum(x[1:ceiling(n*tv[2])] >= qt)/n/tv[2]- tv[1]*tv[2] #tv[2] is v sum(x[1:ceiling(n*tv[2])] >= qt)/n } # compute the correspondence profile # tt, vv: vector between (0, 1) for percentages get.uri.2d <- function(x1, x2, tt, vv, spline.df=NULL){ o <- order(x1, x2, decreasing=T) # sort x2 by the order of x1 x2.ordered <- x2[o] tv <- cbind(tt, vv) ntotal <- length(x1) # number of peaks uri <- apply(tv, 1, comp.uri, x=x2.ordered) # compute the derivative of URI vs t using small bins uri.binned <- uri[seq(1, length(uri), by=4)] tt.binned <- tt[seq(1, length(uri), by=4)] uri.slope <- (uri.binned[2:(length(uri.binned))] - uri.binned[1:(length(uri.binned)-1)])/(tt.binned[2:(length(uri.binned))] - tt.binned[1:(length(tt.binned)-1)]) # smooth uri using spline # first find where the jump is and don't fit the jump # this is the index on the left # jump.left.old <- which.max(uri[-1]-uri[-length(uri)]) short.list.length <- min(sum(x1>0)/length(x1), sum(x2>0)/length(x2)) if(short.list.length < max(tt)){ jump.left <- which(tt>short.list.length)[1]-1 } else { jump.left <- which.max(tt) } # reversed.index <- seq(length(tt), 1, by=-1) # nequal <- sum(uri[reversed.index]== tt[reversed.index]) # temp <- which(uri[reversed.index]== tt[reversed.index])[nequal] # jump.left <- length(tt)-temp if(jump.left < 6){ jump.left <- length(tt) } if(is.null(spline.df)) uri.spl <- smooth.spline(tt[1:jump.left], uri[1:jump.left], df=6.4) else{ uri.spl <- smooth.spline(tt[1:jump.left], uri[1:jump.left], df=spline.df) } # predict the first derivative uri.der <- predict(uri.spl, tt[1:jump.left], deriv=1) invisible(list(tv=tv, uri=uri, uri.slope=uri.slope, t.binned=tt.binned[2:length(uri.binned)], uri.spl=uri.spl, uri.der=uri.der, jump.left=jump.left, ntotal=ntotal)) } # change the scale of uri from based on t (percentage) to n (number of peaks or basepairs) # this is for plotting multiple pairwise URI's on the same plot scale.t2n <- function(uri){ ntotal <- uri$ntotal tv <- uri$tv*uri$ntotal uri.uri <- uri$uri*uri$ntotal jump.left <- uri$jump.left uri.spl <- uri$uri.spl uri.spl$x <- uri$uri.spl$x*uri$ntotal uri.spl$y <- uri$uri.spl$y*uri$ntotal t.binned <- uri$t.binned*uri$ntotal uri.slope <- uri$uri.slope uri.der <- uri$uri.der uri.der$x <- uri$uri.der$x*uri$ntotal uri.der$y <- uri$uri.der$y uri.n <- list(tv=tv, uri=uri.uri, t.binned=t.binned, uri.slope=uri.slope, uri.spl=uri.spl, uri.der=uri.der, ntotal=ntotal, jump.left=jump.left) return(uri.n) } # a wrapper for running URI for peaks from peak calling results # both data1 and data2 are calling results in narrowpeak format compute.pair.uri <- function(data.1, data.2, sig.value1="signal.value", sig.value2="signal.value", spline.df=NULL, overlap.ratio=0){ tt <- seq(0.01, 1, by=0.01) vv <- tt if(sig.value1=="signal.value"){ data.1.enrich <- data.frame(chr=data.1$chr, start.ori=data.1$start.ori, stop.ori=data.1$stop.ori, start=data.1$start, stop=data.1$stop, sig.value=data.1$signal.value, signal.value=data.1$signal.value, p.value=data.1$p.value, q.value=data.1$q.value) } else { if(sig.value1=="p.value"){ data.1.enrich <- data.frame(chr=data.1$chr, start.ori=data.1$start.ori, stop.ori=data.1$stop.ori, start=data.1$start, stop=data.1$stop, sig.value=data.1$p.value, signal.value=data.1$signal.value, p.value=data.1$p.value, q.value=data.1$q.value) } else { if(sig.value1=="q.value"){ data.1.enrich <- data.frame(chr=data.1$chr, start.ori=data.1$start.ori, stop.ori=data.1$stop.ori, start=data.1$start, stop=data.1$stop, sig.value=data.1$q.value, signal.value=data.1$signal.value, p.value=data.1$p.value, q.value=data.1$q.value) } } } if(sig.value2=="signal.value"){ data.2.enrich <- data.frame(chr=data.2$chr, start.ori=data.2$start.ori, stop.ori=data.2$stop.ori, start=data.2$start, stop=data.2$stop, sig.value=data.2$signal.value, signal.value=data.2$signal.value, p.value=data.2$p.value, q.value=data.2$q.value) } else { if(sig.value2=="p.value"){ data.2.enrich <- data.frame(chr=data.2$chr, start.ori=data.2$start.ori, stop.ori=data.2$stop.ori, start=data.2$start, stop=data.2$stop, sig.value=data.2$p.value, signal.value=data.2$signal.value, p.value=data.2$p.value, q.value=data.2$q.value) } else { if(sig.value2=="q.value"){ data.2.enrich <- data.frame(chr=data.2$chr, start.ori=data.2$start.ori, stop.ori=data.2$stop.ori, start=data.2$start, stop=data.2$stop, sig.value=data.2$q.value, signal.value=data.2$signal.value, p.value=data.2$p.value, q.value=data.2$q.value) } } } ### by peaks # data12.enrich <- pair.peaks(data.1.enrich, data.2.enrich) data12.enrich <- pair.peaks.filter(data.1.enrich, data.2.enrich, p.value.impute=0, overlap.ratio) uri <- get.uri.2d(as.numeric(as.character(data12.enrich$merge1$sig.value)), as.numeric(as.character(data12.enrich$merge2$sig.value)), tt, vv, spline.df=spline.df) uri.n <- scale.t2n(uri) return(list(uri=uri, uri.n=uri.n, data12.enrich=data12.enrich, sig.value1=sig.value1, sig.value2=sig.value2)) } # compute uri for matched sample get.uri.matched <- function(data12, df=10){ tt <- seq(0.01, 1, by=0.01) vv <- tt uri <- get.uri.2d(data12$sample1$sig.value, data12$sample2$sig.value, tt, vv, spline.df=df) # change scale from t to n uri.n <- scale.t2n(uri) return(list(uri=uri, uri.n=uri.n)) } # map.uv is a pair of significant values corresponding to specified consistency FDR # assuming values in map.uv and qvalue are linearly related # data.set is the original data set # sig.value is the name of the significant value in map.uv, say enrichment # nominal.value is the one we want to map to, say q-value # map.sig.value <- function(data.set, map.uv, nominal.value){ index.nominal <- which(names(data.set$merge1)==nominal.value) nentry <- nrow(map.uv) map.nominal <- rbind(map.uv[, c("sig.value1", "sig.value2")]) for(i in 1:nentry){ map.nominal[i, "sig.value1"] <- data.set$merge1[unique(which.min(abs(data.set$merge1$sig.value-map.uv[i, "sig.value1"]))), index.nominal] map.nominal[i, "sig.value2"] <- data.set$merge2[unique(which.min(abs(data.set$merge2$sig.value-map.uv[i, "sig.value2"]))), index.nominal] } invisible(map.nominal) } ############### plot correspondence profile # plot multiple comparison wrt one template # uri.list contains the total number of peaks # plot.missing=F: not plot the missing points on the right plot.uri.group <- function(uri.n.list, plot.dir, file.name=NULL, legend.txt, xlab.txt="num of significant peaks", ylab.txt="num of peaks in common", col.start=0, col.txt=NULL, plot.missing=F, title.txt=NULL){ if(is.null(col.txt)) col.txt <- c("black", "red", "purple", "green", "blue", "cyan", "magenta", "orange", "grey") n <- length(uri.n.list) ntotal <- c() for(i in 1:n) ntotal[i] <- uri.n.list[[i]]$ntotal jump.left <- c() jump.left.der <- c() ncommon <- c() for(i in 1:n){ # jump.left[i] <- which.max(uri.n.list[[i]]$uri[-1]-uri.n.list[[i]]$uri[-length(uri.n.list[[i]]$uri)]) # if(jump.left[i] < 6) # jump.left[i] <- length(uri.n.list[[i]]$uri) ## reversed.index <- seq(length(uri.n.list[[i]]$tv[,1]), 1, by=-1) ## nequal <- sum(uri.n.list[[i]]$uri[reversed.index]== uri.n.list[[i]]$tv[reversed.index,1]) ## temp <- which(uri.n.list[[i]]$uri[reversed.index]== uri.n.list[[i]]$tv[reversed.index,1])[nequal] ## jump.left[i] <- length(uri.n.list[[i]]$tv[,1])-temp ##print(uri.n.list[[i]]$uri) ##print(uri.n.list[[i]]$tv[,1]) ## jump.left[i] <- uri.n.list[[i]]$jump.left # jump.left.der[i] <- sum(uri.n.list[[i]]$t.binned < uri.n.list[[i]]$uri.der$x[length(uri.n.list[[i]]$uri.der$x)]) jump.left[i] <- uri.n.list[[i]]$jump.left jump.left.der[i] <- jump.left[i] ncommon[i] <- uri.n.list[[i]]$tv[jump.left[i],1] } if(plot.missing){ max.peak <- max(ntotal) } else { max.peak <- max(ncommon)*1.05 } if(!is.null(file.name)){ postscript(paste(plot.dir, "uri.", file.name, sep="")) par(mfrow=c(1,1), mar=c(5,5,4,2)) } plot(uri.n.list[[1]]$tv[,1], uri.n.list[[1]]$uri, type="n", xlab=xlab.txt, ylab=ylab.txt, xlim=c(0, max.peak), ylim=c(0, max.peak), cex.lab=2) for(i in 1:n){ if(plot.missing){ points(uri.n.list[[i]]$tv[,1], uri.n.list[[i]]$uri, col=col.txt[i+col.start], cex=0.5 ) } else { points(uri.n.list[[i]]$tv[1:jump.left[i],1], uri.n.list[[i]]$uri[1:jump.left[i]], col=col.txt[i+col.start], cex=0.5) } lines(uri.n.list[[i]]$uri.spl, col=col.txt[i+col.start], lwd=4) } abline(coef=c(0,1), lty=3) legend(0, max.peak, legend=legend.txt, col=col.txt[(col.start+1):length(col.txt)], lty=1, lwd=3, cex=2) if(!is.null(title)) title(title.txt) if(!is.null(file.name)){ dev.off() } if(!is.null(file.name)){ postscript(paste(plot.dir, "duri.", file.name, sep="")) par(mfrow=c(1,1), mar=c(5,5,4,2)) } plot(uri.n.list[[1]]$t.binned, uri.n.list[[1]]$uri.slope, type="n", xlab=xlab.txt, ylab="slope", xlim=c(0, max.peak), ylim=c(0, 1.5), cex.lab=2) for(i in 1:n){ # if(plot.missing){ # points(uri.n.list[[i]]$t.binned, uri.n.list[[i]]$uri.slope, col=col.txt[i+col.start], cex=0.5) # } else { # points(uri.n.list[[i]]$t.binned[1:jump.left.der[i]], uri.n.list[[i]]$uri.slope[1:jump.left.der[i]], col=col.txt[i+col.start], cex=0.5) # } lines(uri.n.list[[i]]$uri.der, col=col.txt[i+col.start], lwd=4) } abline(h=1, lty=3) legend(0.5*max.peak, 1.5, legend=legend.txt, col=col.txt[(col.start+1):length(col.txt)], lty=1, lwd=3, cex=2) if(!is.null(title)) title(title.txt) if(!is.null(file.name)){ dev.off() } } ####################### ####################### copula fitting for matched peaks ####################### # estimation from mixed copula model # 4-5-09 # A nonparametric estimation of mixed copula model # updated # c1, c2, f1, f2, g1, g2 are vectors # c1*f1*g1 and c2*f2*g2 are copula densities for the two components # xd1 and yd1 are the values of marginals for the first component # xd2 and yd2 are the values of marginals for the 2nd component # # ez is the prob for being in the consistent group get.ez <- function(p, c1, c2, xd1, yd1, xd2, yd2){ return(p*c1*xd1*yd1/(p*c1*xd1*yd1 + (1-p)*c2*xd2*yd2)) } # checked # this is C_12 not the copula density function c=C_12 * f1* f2 # since nonparametric estimation is used here for f1 and f2, which # are constant throughout the iterations, we don't need them for optimization # # bivariate gaussian copula function # t and s are vectors of same length, both are percentiles # return a vector gaussian.cop.den <- function(t, s, rho){ A <- qnorm(t)^2 + qnorm(s)^2 B <- qnorm(t)*qnorm(s) loglik <- -log(1-rho^2)/2 - rho/(2*(1-rho^2))*(rho*A-2*B) return(exp(loglik)) } clayton.cop.den <- function(t, s, rho){ if(rho > 0) return(exp(log(rho+1)-(rho+1)*(log(t)+log(s))-(2+1/rho)*log(t^(-rho) + s^(-rho)-1))) if(rho==0) return(1) if(rho<0) stop("Incorrect Clayton copula coefficient") } # checked # estimate rho from Gaussian copula mle.gaussian.copula <- function(t, s, e.z){ # reparameterize to bound from rho=+-1 l.c <- function(rho, t, s, e.z){ # cat("rho=", rho, "\n") sum(e.z*log(gaussian.cop.den(t, s, rho)))} rho.max <- optimize(f=l.c, c(-0.998, 0.998), maximum=T, tol=0.00001, t=t, s=s, e.z=e.z) #print(rho.max$m) #cat("cor=", cor(qnorm(t)*e.z, qnorm(s)*e.z), "\t", "rho.max=", rho.max$m, "\n") # return(sign(rho.max$m)/(1+rho.max$m)) return(rho.max$m) } # estimate mle from Clayton copula, mle.clayton.copula <- function(t, s, e.z){ l.c <- function(rho, t, s, e.z){ lc <- sum(e.z*log(clayton.cop.den(t, s, rho))) # cat("rho=", rho, "\t", "l.c=", lc, "\n") return(lc) } rho.max <- optimize(f=l.c, c(0.1, 20), maximum=T, tol=0.00001, t=t, s=s, e.z=e.z) return(rho.max$m) } # updated # mixture likelihood of two gaussian copula # nonparametric and ranked transformed loglik.2gaussian.copula <- function(x, y, p, rho1, rho2, x.mar, y.mar){ px.1 <- get.pdf.cdf(x, x.mar$f1) px.2 <- get.pdf.cdf(x, x.mar$f2) py.1 <- get.pdf.cdf(y, y.mar$f1) py.2 <- get.pdf.cdf(y, y.mar$f2) c1 <- gaussian.cop.den(px.1$cdf, py.1$cdf, rho1) c2 <- gaussian.cop.den(px.2$cdf, py.2$cdf, rho2) sum(log(p*c1*px.1$pdf*py.1$pdf + (1-p)*c2*px.2$pdf*py.2$pdf)) } loglik.2copula <- function(x, y, p, rho1, rho2, x.mar, y.mar, copula.txt){ px.1 <- pdf.cdf$px.1 px.2 <- pdf.cdf$px.2 py.1 <- pdf.cdf$py.1 py.2 <- pdf.cdf$py.2 if(copula.txt=="gaussian"){ c1 <- gaussian.cop.den(px.1$cdf, py.1$cdf, rho1) c2 <- gaussian.cop.den(px.2$cdf, py.2$cdf, rho2) } else { if(copula.txt=="clayton"){ c1 <- clayton.cop.den(px.1$cdf, py.1$cdf, rho1) c2 <- clayton.cop.den(px.2$cdf, py.2$cdf, rho2) } } sum(log(p*c1*px.1$pdf*py.1$pdf + (1-p)*c2*px.2$pdf*py.2$pdf)) } # estimate the marginals of each component using histogram estimator in EM # return the density, breaks, and cdf of the histogram estimator est.mar.hist <- function(x, e.z, breaks){ binwidth <- c() nbin <- length(breaks)-1 nx <- length(x) # the histogram x1.pdf <- c() x2.pdf <- c() x1.cdf <- c() x2.cdf <- c() # the pdf for each point x1.pdf.value <- rep(NA, nx) x2.pdf.value <- rep(NA, nx) x1.cdf.value <- rep(NA, nx) x2.cdf.value <- rep(NA, nx) for(i in 1:nbin){ binwidth[i] <- breaks[i+1] - breaks[i] if(i < nbin) in.bin <- x>= breaks[i] & x < breaks[i+1] else # last bin in.bin <- x>= breaks[i] & x <=breaks[i+1] # each bin add one observation to avoid empty bins # multiple (nx+nbin)/(nx+nbin+1) to avoid blowup when looking up for # quantiles x1.pdf[i] <- (sum(e.z[in.bin])+1)/(sum(e.z)+nbin)/binwidth[i]*(nx+nbin)/(nx+nbin+1) x2.pdf[i] <- (sum(1-e.z[in.bin])+1)/(sum(1-e.z)+nbin)/binwidth[i]*(nx+nbin)/(nx+nbin+1) # x1.pdf[i] <- sum(e.z[in.bin])/sum(e.z)/binwidth[i]*nx/(nx+1) # x2.pdf[i] <- sum(1-e.z[in.bin])/sum(1-e.z)/binwidth[i]*nx/(nx+1) # treat each bin as a value for a discrete variable # x1.cdf[i] <- sum(x1.pdf[1:i]*binwidth[1:i]) # x2.cdf[i] <- sum(x2.pdf[1:i]*binwidth[1:i]) # cumulative density before reaching i if(i>1){ x1.cdf[i] <- sum(x1.pdf[1:(i-1)]*binwidth[1:(i-1)]) x2.cdf[i] <- sum(x2.pdf[1:(i-1)]*binwidth[1:(i-1)]) } else{ x1.cdf[i] <- 0 x2.cdf[i] <- 0 } # make a vector of nx to store the values of pdf and cdf for each x # this will speed up the computation dramatically x1.pdf.value[in.bin] <- x1.pdf[i] x2.pdf.value[in.bin] <- x2.pdf[i] x1.cdf.value[in.bin] <- x1.cdf[i] + x1.pdf[i]*(x[in.bin]-breaks[i]) x2.cdf.value[in.bin] <- x2.cdf[i] + x2.pdf[i]*(x[in.bin]-breaks[i]) } # x1.cdf <- cumsum(x1.pdf*binwidth) # x2.cdf <- cumsum(x2.pdf*binwidth) f1 <-list(breaks=breaks, density=x1.pdf, cdf=x1.cdf) f2 <-list(breaks=breaks, density=x2.pdf, cdf=x2.cdf) f1.value <- list(pdf=x1.pdf.value, cdf=x1.cdf.value) f2.value <- list(pdf=x2.pdf.value, cdf=x2.cdf.value) return(list(f1=f1, f2=f2, f1.value=f1.value, f2.value=f2.value)) } # estimate the marginal cdf from rank est.cdf.rank <- function(x, conf.z){ # add 1 to prevent blow up x1.cdf <- rank(x[conf.z==1])/(length(x[conf.z==1])+1) x2.cdf <- rank(x[conf.z==0])/(length(x[conf.z==0])+1) return(list(cdf1=x1.cdf, cdf2=x2.cdf)) } # df is a density function with fields: density, cdf and breaks, x is a scalar get.pdf <- function(x, df){ if(x < df$breaks[1]) cat("x is out of the range of df\n") index <- which(df$breaks >= x)[1] if(index==1) index <- index +1 return(df$density[index-1]) } # get cdf from histgram estimator for a single value get.cdf <- function(x, df){ index <- which(df$breaks >= x)[1] if(index==1) index <- index +1 return(df$cdf[index-1]) } # df is a density function with fields: density, cdf and breaks get.pdf.cdf <- function(x.vec, df){ x.pdf <- sapply(x.vec, get.pdf, df=df) x.cdf <- sapply(x.vec, get.cdf, df=df) return(list(cdf=x.cdf, pdf=x.pdf)) } # E-step # x and y are the original observations or ranks # rho1 and rho2 are the parameters of each copula # f1, f2, g1, g2 are functions, each is a histogram e.step.2gaussian <- function(x, y, p, rho1, rho2, x.mar, y.mar){ # get pdf and cdf of each component from functions in the corresponding component px.1 <- get.pdf.cdf(x, x.mar$f1) px.2 <- get.pdf.cdf(x, x.mar$f2) py.1 <- get.pdf.cdf(y, y.mar$f1) py.2 <- get.pdf.cdf(y, y.mar$f2) c1 <- gaussian.cop.den(px.1$cdf, py.1$cdf, rho1) c2 <- gaussian.cop.den(px.2$cdf, py.2$cdf, rho2) return(get.ez(p, c1, c2, px.1$pdf, py.1$pdf, px.2$pdf, py.2$pdf)) } # E-step # rho1 and rho2 are the parameters of each copula e.step.2copula <- function(x, y, p, rho1, rho2, x.mar, y.mar, copula.txt){ # get pdf and cdf of each component from functions in the corresponding component px.1 <- get.pdf.cdf(x, x.mar$f1) px.2 <- get.pdf.cdf(x, x.mar$f2) py.1 <- get.pdf.cdf(y, y.mar$f1) py.2 <- get.pdf.cdf(y, y.mar$f2) if(copula.txt=="gaussian"){ c1 <- gaussian.cop.den(px.1$cdf, py.1$cdf, rho1) c2 <- gaussian.cop.den(px.2$cdf, py.2$cdf, rho2) } else { if(copula.txt=="clayton"){ c1 <- clayton.cop.den(px.1$cdf, py.1$cdf, rho1) c2 <- clayton.cop.den(px.2$cdf, py.2$cdf, rho2) } } return(get.ez(p, c1, c2, px.1$pdf, py.1$pdf, px.2$pdf, py.2$pdf)) } # M-step m.step.2gaussian <- function(x, y, e.z, breaks){ # compute f1, f2, g1 and g2 x.mar <- est.mar.hist(x, e.z, breaks) y.mar <- est.mar.hist(y, e.z, breaks) px.1 <- get.pdf.cdf(x, x.mar$f1) px.2 <- get.pdf.cdf(x, x.mar$f2) py.1 <- get.pdf.cdf(y, y.mar$f1) py.2 <- get.pdf.cdf(y, y.mar$f2) rho1 <- mle.gaussian.copula(px.1$cdf, py.1$cdf, e.z) rho2 <- mle.gaussian.copula(px.2$cdf, py.2$cdf, 1-e.z) p <- sum(e.z)/length(e.z) return(list(p=p, rho1=rho1, rho2=rho2, x.mar=x.mar, y.mar=y.mar)) } m.step.2copula <- function(x, y, e.z, breaks, copula.txt){ # compute f1, f2, g1 and g2 x.mar <- est.mar.hist(x, e.z, breaks) y.mar <- est.mar.hist(y, e.z, breaks) px.1 <- get.pdf.cdf(x, x.mar$f1) px.2 <- get.pdf.cdf(x, x.mar$f2) py.1 <- get.pdf.cdf(y, y.mar$f1) py.2 <- get.pdf.cdf(y, y.mar$f2) if(copula.txt=="gaussian"){ rho1 <- mle.gaussian.copula(px.1$cdf, py.1$cdf, e.z) rho2 <- mle.gaussian.copula(px.2$cdf, py.2$cdf, 1-e.z) } else { if(copula.txt=="clayton"){ rho1 <- mle.clayton.copula(px.1$cdf, py.1$cdf, e.z) rho2 <- mle.clayton.copula(px.2$cdf, py.2$cdf, 1-e.z) } } p <- sum(e.z)/length(e.z) return(list(p=p, rho1=rho1, rho2=rho2, x.mar=x.mar, y.mar=y.mar)) } # E-step: pass values # x and y are the original observations or ranks # rho1 and rho2 are the parameters of each copula # f1, f2, g1, g2 are functions, each is a histogram e.step.2gaussian.value <- function(x, y, p, rho1, rho2, pdf.cdf){ c1 <- gaussian.cop.den(pdf.cdf$px.1$cdf, pdf.cdf$py.1$cdf, rho1) c2 <- gaussian.cop.den(pdf.cdf$px.2$cdf, pdf.cdf$py.2$cdf, rho2) e.z <- get.ez(p, c1, c2, pdf.cdf$px.1$pdf, pdf.cdf$py.1$pdf, pdf.cdf$px.2$pdf, pdf.cdf$py.2$pdf) return(e.z) } e.step.2copula.value <- function(x, y, p, rho1, rho2, pdf.cdf, copula.txt){ if(copula.txt =="gaussian"){ c1 <- gaussian.cop.den(pdf.cdf$px.1$cdf, pdf.cdf$py.1$cdf, rho1) c2 <- gaussian.cop.den(pdf.cdf$px.2$cdf, pdf.cdf$py.2$cdf, rho2) } else { if(copula.txt =="clayton"){ c1 <- clayton.cop.den(pdf.cdf$px.1$cdf, pdf.cdf$py.1$cdf, rho1) c2 <- clayton.cop.den(pdf.cdf$px.2$cdf, pdf.cdf$py.2$cdf, rho2) } } e.z <- get.ez(p, c1, c2, pdf.cdf$px.1$pdf, pdf.cdf$py.1$pdf, pdf.cdf$px.2$pdf, pdf.cdf$py.2$pdf) return(e.z) } # M-step: pass values m.step.2gaussian.value <- function(x, y, e.z, breaks, fix.rho2){ # compute f1, f2, g1 and g2 x.mar <- est.mar.hist(x, e.z, breaks) y.mar <- est.mar.hist(y, e.z, breaks) # px.1 <- get.pdf.cdf(x, x.mar$f1) # px.2 <- get.pdf.cdf(x, x.mar$f2) # py.1 <- get.pdf.cdf(y, y.mar$f1) # py.2 <- get.pdf.cdf(y, y.mar$f2) px.1 <- x.mar$f1.value px.2 <- x.mar$f2.value py.1 <- y.mar$f1.value py.2 <- y.mar$f2.value rho1 <- mle.gaussian.copula(px.1$cdf, py.1$cdf, e.z) if(!fix.rho2) rho2 <- mle.gaussian.copula(px.2$cdf, py.2$cdf, 1-e.z) else rho2 <- 0 p <- sum(e.z)/length(e.z) pdf.cdf <- list(px.1=px.1, px.2=px.2, py.1=py.1, py.2=py.2) return(list(p=p, rho1=rho1, rho2=rho2, x.mar=x.mar, y.mar=y.mar, pdf.cdf=pdf.cdf)) } m.step.2gaussian.value2 <- function(x, y, e.z, breaks, fix.rho2, x.mar, y.mar){ # compute f1, f2, g1 and g2 # x.mar <- est.mar.hist(x, e.z, breaks) # y.mar <- est.mar.hist(y, e.z, breaks) # px.1 <- get.pdf.cdf(x, x.mar$f1) # px.2 <- get.pdf.cdf(x, x.mar$f2) # py.1 <- get.pdf.cdf(y, y.mar$f1) # py.2 <- get.pdf.cdf(y, y.mar$f2) px.1 <- x.mar$f1.value px.2 <- x.mar$f2.value py.1 <- y.mar$f1.value py.2 <- y.mar$f2.value rho1 <- mle.gaussian.copula(px.1$cdf, py.1$cdf, e.z) if(!fix.rho2) rho2 <- mle.gaussian.copula(px.2$cdf, py.2$cdf, 1-e.z) else rho2 <- 0 p <- sum(e.z)/length(e.z) pdf.cdf <- list(px.1=px.1, px.2=px.2, py.1=py.1, py.2=py.2) return(list(p=p, rho1=rho1, rho2=rho2, x.mar=x.mar, y.mar=y.mar, pdf.cdf=pdf.cdf)) } m.step.2copula.value <- function(x, y, e.z, breaks, fix.rho2, copula.txt){ # compute f1, f2, g1 and g2 x.mar <- est.mar.hist(x, e.z, breaks) y.mar <- est.mar.hist(y, e.z, breaks) # px.1 <- get.pdf.cdf(x, x.mar$f1) # px.2 <- get.pdf.cdf(x, x.mar$f2) # py.1 <- get.pdf.cdf(y, y.mar$f1) # py.2 <- get.pdf.cdf(y, y.mar$f2) px.1 <- x.mar$f1.value px.2 <- x.mar$f2.value py.1 <- y.mar$f1.value py.2 <- y.mar$f2.value if(copula.txt=="gaussian"){ rho1 <- mle.gaussian.copula(px.1$cdf, py.1$cdf, e.z) if(!fix.rho2) rho2 <- mle.gaussian.copula(px.2$cdf, py.2$cdf, 1-e.z) else rho2 <- 0 } else { if(copula.txt=="clayton"){ rho1 <- mle.clayton.copula(px.1$cdf, py.1$cdf, e.z) if(!fix.rho2) rho2 <- mle.clayton.copula(px.2$cdf, py.2$cdf, 1-e.z) else rho2 <- 0 } } p <- sum(e.z)/length(e.z) pdf.cdf <- list(px.1=px.1, px.2=px.2, py.1=py.1, py.2=py.2) return(list(p=p, rho1=rho1, rho2=rho2, x.mar=x.mar, y.mar=y.mar, pdf.cdf=pdf.cdf)) } # updated # mixture likelihood of two gaussian copula # nonparametric and ranked transformed loglik.2gaussian.copula.value <- function(x, y, p, rho1, rho2, pdf.cdf){ px.1 <- pdf.cdf$px.1 px.2 <- pdf.cdf$px.2 py.1 <- pdf.cdf$py.1 py.2 <- pdf.cdf$py.2 c1 <- gaussian.cop.den(px.1$cdf, py.1$cdf, rho1) c2 <- gaussian.cop.den(px.2$cdf, py.2$cdf, rho2) sum(log(p*c1*px.1$pdf*py.1$pdf + (1-p)*c2*px.2$pdf*py.2$pdf)) } # updated # mixture likelihood of two gaussian copula # nonparametric and ranked transformed loglik.2copula.value <- function(x, y, p, rho1, rho2, pdf.cdf, copula.txt){ px.1 <- pdf.cdf$px.1 px.2 <- pdf.cdf$px.2 py.1 <- pdf.cdf$py.1 py.2 <- pdf.cdf$py.2 if(copula.txt=="gaussian"){ c1 <- gaussian.cop.den(px.1$cdf, py.1$cdf, rho1) c2 <- gaussian.cop.den(px.2$cdf, py.2$cdf, rho2) } else { if(copula.txt=="clayton"){ c1 <- clayton.cop.den(px.1$cdf, py.1$cdf, rho1) c2 <- clayton.cop.den(px.2$cdf, py.2$cdf, rho2) } } sum(log(p*c1*px.1$pdf*py.1$pdf + (1-p)*c2*px.2$pdf*py.2$pdf)) } # EM for 2 Gaussian, speed up computation, unfinished em.2gaussian.quick <- function(x, y, p0, rho1.0, rho2.0, eps, fix.p=F, stoc=T, fix.rho2=T){ x <- rank(x, tie="random") y <- rank(y, tie="random") # x <- rank(x, tie="average") # y <- rank(y, tie="average") # nbin=20 xy.min <- min(x, y) xy.max <- max(x, y) binwidth <- (xy.max-xy.min)/50 breaks <- seq(xy.min-binwidth/100, xy.max+binwidth/100, by=(xy.max-xy.min+binwidth/50)/50) # breaks <- seq(xy.min, xy.max, by=binwidth) # initiate marginals # initialization: first p0 data has # e.z <- e.step.2gaussian(x, y, p0, rho1.0, rho2.0, x0.mar, y0.mar) # this starting point assumes two components are overlapped e.z <- c(rep(0.9, round(length(x)*p0)), rep(0.1, length(x)-round(length(x)*p0))) if(!stoc) para <- m.step.2gaussian.value(x, y, e.z, breaks, fix.rho2) else para <- m.step.2gaussian.stoc.value(x, y, e.z, breaks, fix.rho2) if(fix.p){ p <- p0 } else { p <- para$p } if(fix.rho2){ rho2 <- rho2.0 } else { rho2 <- para$rho2 } # rho1 <- 0.8 rho1 <- para$rho1 l0 <- loglik.2gaussian.copula.value(x, y, p, rho1, rho2, para$pdf.cdf) loglik.trace <- c() loglik.trace[1] <- l0 # loglik.trace[2] <- l1 to.run <- T i <- 2 # this two lines to remove # x.mar <- est.mar.hist(x, e.z, breaks) # y.mar <- est.mar.hist(y, e.z, breaks) while(to.run){ e.z <- e.step.2gaussian.value(x, y, p, rho1, rho2, para$pdf.cdf) if(!stoc) para <- m.step.2gaussian.value(x, y, e.z, breaks, fix.rho2) else para <- m.step.2gaussian.stoc.value(x, y, e.z, breaks, fix.rho2) # fix x.mar and y.mar : to remove # if(!stoc) # para <- m.step.2gaussian.value2(x, y, e.z, breaks, fix.rho2, x.mar, y.mar) # else # para <- m.step.2gaussian.stoc.value(x, y, e.z, breaks, fix.rho2) if(fix.p){ p <- p0 } else { p <- para$p } if(fix.rho2){ rho2 <- rho2.0 } else { rho2 <- para$rho2 } # rho1 <- 0.8 rho1 <- para$rho1 # l0 <- l1 l1 <- loglik.2gaussian.copula.value(x, y, p, rho1, rho2, para$pdf.cdf) loglik.trace[i] <- l1 #cat("l1=", l1, "\n") # Aitken acceleration criterion if(i > 2){ l.inf <- loglik.trace[i-2] + (loglik.trace[i-1] - loglik.trace[i-2])/(1-(loglik.trace[i]-loglik.trace[i-1])/(loglik.trace[i-1]-loglik.trace[i-2])) to.run <- abs(l.inf - loglik.trace[i]) > eps #cat("para=", "p=", para$p, " rho1=", rho1, " rho2=", rho2, "\n") #cat("l.inf=", l.inf, "\n") #cat(l.inf-loglik.trace[i], "\n") } i <- i+1 } bic <- -2*l1 + (2*(length(breaks)-1+1)+1-fix.p-fix.rho2)*log(length(x)) # parameters return(list(para=list(p=para$p, rho1=rho1, rho2=rho2), loglik=l1, bic=bic, e.z=e.z, conf.z = para$conf.z, loglik.trace=loglik.trace, x.mar=para$x.mar, y.mar=para$y.mar, breaks=breaks)) } em.2copula.quick <- function(x, y, p0, rho1.0, rho2.0, eps, fix.p=F, stoc=T, fix.rho2=T, copula.txt, nbin=50){ x <- rank(x, tie="random") y <- rank(y, tie="random") # x <- rank(x, tie="first") # y <- rank(y, tie="first") # nbin=50 xy.min <- min(x, y) xy.max <- max(x, y) binwidth <- (xy.max-xy.min)/50 breaks <- seq(xy.min-binwidth/100, xy.max+binwidth/100, by=(xy.max-xy.min+binwidth/50)/nbin) # breaks <- seq(xy.min, xy.max, by=binwidth) # initiate marginals # initialization: first p0 data has # e.z <- e.step.2gaussian(x, y, p0, rho1.0, rho2.0, x0.mar, y0.mar) # this starting point assumes two components are overlapped e.z <- c(rep(0.9, round(length(x)*p0)), rep(0.1, length(x)-round(length(x)*p0))) if(!stoc) para <- m.step.2copula.value(x, y, e.z, breaks, fix.rho2, copula.txt) else para <- m.step.2copula.stoc.value(x, y, e.z, breaks, fix.rho2, copula.txt) if(fix.p){ p <- p0 } else { p <- para$p } if(fix.rho2){ rho2 <- rho2.0 } else { rho2 <- para$rho2 } l0 <- loglik.2copula.value(x, y, p, para$rho1, rho2, para$pdf.cdf, copula.txt) loglik.trace <- c() loglik.trace[1] <- l0 # loglik.trace[2] <- l1 to.run <- T i <- 2 while(to.run){ e.z <- e.step.2copula.value(x, y, p, para$rho1, rho2, para$pdf.cdf, copula.txt) if(!stoc) para <- m.step.2copula.value(x, y, e.z, breaks, fix.rho2, copula.txt) else para <- m.step.2copula.stoc.value(x, y, e.z, breaks, fix.rho2, copula.txt) if(fix.p){ p <- p0 } else { p <- para$p } if(fix.rho2){ rho2 <- rho2.0 } else { rho2 <- para$rho2 } # l0 <- l1 l1 <- loglik.2copula.value(x, y, p, para$rho1, rho2, para$pdf.cdf, copula.txt) loglik.trace[i] <- l1 cat("l1=", l1, "\n") # Aitken acceleration criterion if(i > 2){ l.inf <- loglik.trace[i-2] + (loglik.trace[i-1] - loglik.trace[i-2])/(1-(loglik.trace[i]-loglik.trace[i-1])/(loglik.trace[i-1]-loglik.trace[i-2])) to.run <- abs(l.inf - loglik.trace[i]) > eps cat("para=", "p=", para$p, " rho1=", para$rho1, " rho2=", rho2, "\n") #cat("l.inf=", l.inf, "\n") #cat(l.inf-loglik.trace[i], "\n") } i <- i+1 } bic <- -2*l1 + (2*(length(breaks)-1+1)+1-fix.p-fix.rho2)*log(length(x)) # parameters return(list(para=list(p=para$p, rho1=para$rho1, rho2=rho2), loglik=l1, bic=bic, e.z=e.z, conf.z = para$conf.z, loglik.trace=loglik.trace, x.mar=para$x.mar, y.mar=para$y.mar, breaks=breaks)) } ####################### ####################### fit EM procedure for the matched peaks ####################### # remove the unmatched ones #rm.unmatch <- function(sample1, sample2, p.value.impute=0){ # # sample1.prune <- sample1[sample1$sig.value > p.value.impute & sample2$sig.value > p.value.impute,] # sample2.prune <- sample2[sample1$sig.value > p.value.impute & sample2$sig.value > p.value.impute,] # # invisible(list(sample1=sample1.prune$sig.value, sample2=sample2.prune$sig.value)) #} # fit 2-component model #fit.em <- function(sample12, fix.rho2=T){ # # prune.sample <- rm.unmatch(sample12$merge1, sample12$merge2) # # em.fit <- em.2gaussian.quick(-prune.sample$sample1, -prune.sample$sample2, # p0=0.5, rho1.0=0.7, rho2.0=0, eps=0.01, fix.p=F, stoc=F, fix.rho2) # # invisible(list(em.fit=em.fit, data.pruned=prune.sample)) #} rm.unmatch <- function(sample1, sample2, p.value.impute=0){ sample1.prune <- sample1[sample1$sig.value > p.value.impute & sample2$sig.value > p.value.impute,] sample2.prune <- sample2[sample1$sig.value > p.value.impute & sample2$sig.value > p.value.impute,] invisible(list(sample1=sample1.prune, sample2=sample2.prune)) } # fit 2-component model fit.em <- function(sample12, fix.rho2=T){ prune.sample <- rm.unmatch(sample12$merge1, sample12$merge2) em.fit <- em.2gaussian.quick(-prune.sample$sample1$sig.value, -prune.sample$sample2$sig.value, p0=0.5, rho1.0=0.7, rho2.0=0, eps=0.01, fix.p=F, stoc=F, fix.rho2) invisible(list(em.fit=em.fit, data.pruned=prune.sample)) } fit.2copula.em <- function(sample12, fix.rho2=T, copula.txt){ prune.sample <- rm.unmatch(sample12$merge1, sample12$merge2) # o <- order(prune.sample$sample1) # n <- length(prune.sample$sample1) # para <- init(prune.sample$sample1$sig.value, prune.sample$sample2$sig.value, c(rep(0, round(n/3)), rep(c(0,1), round(n/6)), rep(1, n-round(n/3)-round(n/6)))) # temp <- init.dist(f0, f1) para <- list() para$rho <- 0.6 para$p <- 0.3 para$mu <- 2.5 para$sigma <- 1 ## para$mu <- -temp$mu ## para$sigma <- temp$sigma #cat("mu=", para$mu, "sigma=", para$sigma, "\n") # em.fit <- em.transform.1loop(-prune.sample$sample1, -prune.sample$sample2, cat("EM is running") em.fit <- em.transform(prune.sample$sample1$sig.value, prune.sample$sample2$sig.value, para$mu, para$sigma, para$rho, para$p, eps=0.01) invisible(list(em.fit=em.fit, data.pruned=prune.sample)) } # fit 1-component model fit.1.component <- function(data.pruned, breaks){ # gaussian.1 <- fit.gaussian.1(-data.pruned$sample1$sig.value, -data.pruned$sample2$sig.value, breaks) # clayton.1 <- fit.clayton.1(-data.pruned$sample1$sig.value, -data.pruned$sample2$sig.value, breaks) gaussian.1 <- fit.gaussian.1(-data.pruned$sample1, -data.pruned$sample2, breaks) clayton.1 <- fit.clayton.1(-data.pruned$sample1, -data.pruned$sample2, breaks) return(list(gaussian.1=gaussian.1, clayton.1=clayton.1)) } ################# # Fit a single component ################# # a single gaussian copula # if breaks=NULL, use empirical pdf, otherwise use histogram estimate fit.gaussian.1 <- function(x, y, breaks=NULL){ # rank transformed and compute the empirical cdf t <- emp.mar.cdf.rank(x) s <- emp.mar.cdf.rank(y) mle.rho <- mle.gaussian.copula(t, s, rep(1, length(t))) c1 <- gaussian.cop.den(t, s, mle.rho) cat("c1", sum(log(c1)), "\n") if(is.null(breaks)){ f1 <- emp.mar.pdf.rank(t) f2 <- emp.mar.pdf.rank(s) } else { x.mar <- est.mar.hist(rank(x), rep(1, length(x)), breaks) y.mar <- est.mar.hist(rank(y), rep(1, length(y)), breaks) f1 <- x.mar$f1.value$pdf # only one component f2 <- y.mar$f1.value$pdf } cat("f1", sum(log(f1)), "\n") cat("f2", sum(log(f2)), "\n") loglik <- sum(log(c1)+log(f1)+log(f2)) bic <- -2*loglik + log(length(t))*(1+length(breaks)-1) return(list(rho=mle.rho, loglik=loglik, bic=bic)) } # a single Clayton copula fit.clayton.1 <- function(x, y, breaks=NULL){ # rank transformed and compute the empirical cdf t <- emp.mar.cdf.rank(x) s <- emp.mar.cdf.rank(y) mle.rho <- mle.clayton.copula(t, s, rep(1, length(t))) c1 <- clayton.cop.den(t, s, mle.rho) if(is.null(breaks)){ f1 <- emp.mar.pdf.rank(t) f2 <- emp.mar.pdf.rank(s) } else { x.mar <- est.mar.hist(rank(x), rep(1, length(x)), breaks) y.mar <- est.mar.hist(rank(y), rep(1, length(y)), breaks) f1 <- x.mar$f1.value$pdf # only one component f2 <- y.mar$f1.value$pdf } loglik <- sum(log(c1)+log(f1)+log(f2)) bic <- -2*loglik + log(length(t))*(1+length(breaks)-1) return(list(rho=mle.rho, tau=rho/(rho+2), loglik=loglik, bic=bic)) } ## obsolete function (01-06-2010) ## compute the average posterior probability to belong to the random component ## for peaks selected at different cutoffs comp.uri.ez <- function(tt, u, v, e.z){ u.t <- quantile(u, prob=(1-tt)) v.t <- quantile(v, prob=(1-tt)) # ez <- mean(e.z[u >= u.t & v >=u.t]) Is this wrong? ez <- mean(e.z[u >= u.t & v >=v.t]) return(ez) } ## obsolete function (01-06-2010) # compute the largest posterior error probability corresponding to # the square centered at the origin and spanned top tt% on both coordinates # so the consistent low rank ones are excluded # boundary.txt: either "max" or "min", if it is error prob, use "max" comp.ez.cutoff <- function(tt, u, v, e.z, boundary.txt){ u.t <- quantile(u, prob=(1-tt)) v.t <- quantile(v, prob=(1-tt)) if(boundary.txt == "max"){ # ez.bound <- max(e.z[u >= u.t & v >=u.t]) ez.bound <- max(e.z[u >= u.t & v >=v.t]) } else { # ez.bound <- min(e.z[u >= u.t & v >=u.t]) ez.bound <- min(e.z[u >= u.t & v >=v.t]) } return(ez.bound) } # obsolete functions: 01-06-2010 # compute the error rate # u.t and v.t are the quantiles # this one is used for the plots generated initially in the brief writeup # and it was used for processing merged data in July before the IDR definition # is formalized # It does not implement the current definition of IDR get.ez.tt.old <- function(em.fit, reverse=T, fdr.level=c(0.01, 0.05, 0.1)){ u <- em.fit$data.pruned$sample1 v <- em.fit$data.pruned$sample2 tt <- seq(0.01, 0.99, by=0.01) if(reverse){ e.z <- 1-em.fit$em.fit$e.z # this is the error prob uri.ez <- sapply(tt, comp.uri.ez, u=u, v=v, e.z=e.z) ez.bound <- sapply(tt, comp.ez.cutoff, u=u, v=v, e.z=e.z, boundary.txt="max") } else { e.z <- em.fit$em.fit$e.z uri.ez <- sapply(tt, comp.uri.ez, u=u, v=v, e.z=e.z) ez.bound <- sapply(tt, comp.ez.cutoff, u=u, v=v, e.z=e.z, boundary.txt="min") } u.t <- quantile(u, prob=(1-tt)) v.t <- quantile(v, prob=(1-tt)) # find the levels on the two replicates sig.value1 <- c() sig.value2 <- c() error.prob.cutoff <- c() n.selected.match <- c() for(i in 1:length(fdr.level)){ # find which uri.ez is closet to fdr.level index <- which.min(abs(uri.ez - fdr.level[i])) sig.value1[i] <- u.t[index] sig.value2[i] <- v.t[index] error.prob.cutoff[i] <- ez.bound[index] if(reverse){ n.selected.match[i] <- sum(e.z<=ez.bound[index]) } else { n.selected.match[i] <- sum(e.z>=ez.bound[index]) } } # output the cutoff of posterior probability, signal values on two replicates map.uv <- cbind(error.prob.cutoff, sig.value1, sig.value2, n.selected.match) return(list(n=tt*length(u), uri.ez=uri.ez, u.t=u.t, v.t=v.t, tt=tt, fdr.level=fdr.level, map.uv=map.uv, e.z=e.z, error.prob.cutoff=error.prob.cutoff)) } # created: 01-06-2010 # Output IDR at various number of selected peaks # Find cutoff (idr cutoff, sig.value cutoff on each replicate) for specified IDR level # IDR definition is similar to FDR get.ez.tt <- function(em.fit, idr.level=c(0.01, 0.05, 0.1)){ # u <- em.fit$data.pruned$sample1$sig.value # v <- em.fit$data.pruned$sample2$sig.value u <- em.fit$data.pruned$sample1 v <- em.fit$data.pruned$sample2 e.z <- 1-em.fit$em.fit$e.z # this is the error prob o <- order(e.z) e.z.ordered <- e.z[o] n.select <- c(1:length(e.z)) IDR <- cumsum(e.z.ordered)/n.select u.o <- u[o] v.o <- v[o] n.level <- length(idr.level) # sig.value1 <- rep(NA, n.level) # sig.value2 <- rep(NA, n.level) ez.cutoff <- rep(NA, n.level) n.selected <- rep(NA, n.level) for(i in 1:length(idr.level)){ # find which uri.ez is closet to fdr.level index <- which.min(abs(IDR - idr.level[i])) # sig.value1[i] <- min(u.o[1:index]) # sig.value2[i] <- min(v.o[1:index]) ez.cutoff[i] <- e.z[index] n.selected[i] <- sum(e.z<=ez.cutoff[i]) } # output the cutoff of posterior probability, number of selected overlapped peaks # map.uv <- cbind(ez.cutoff, sig.value1, sig.value2, n.selected) map.uv <- cbind(ez.cutoff, n.selected) return(list(n=n.select, IDR=IDR, idr.level=idr.level, map.uv=map.uv)) } # return(list(n=tt*length(u), uri.ez=uri.ez, fdr.level=fdr.level, map.uv=map.uv, e.z=e.z, error.prob.cutoff=error.prob.cutoff)) ### compute the mean of the marginals get.mar.mean <- function(em.out){ x.f1 <- em.out$x.mar$f1 x.f2 <- em.out$x.mar$f2 y.f1 <- em.out$y.mar$f1 y.f2 <- em.out$y.mar$f2 x.stat1 <- get.hist.mean(x.f1) x.stat2 <- get.hist.mean(x.f2) y.stat1 <- get.hist.mean(y.f1) y.stat2 <- get.hist.mean(y.f2) return(list(x.mean1=x.stat1$mean, x.mean2=x.stat2$mean, y.mean1=y.stat1$mean, y.mean2=y.stat2$mean, x.sd1=x.stat1$sd, x.sd2=x.stat2$sd, y.sd1=y.stat1$sd, y.sd2=y.stat2$sd )) } # compute the mean of marginals get.hist.mean <- function(x.f){ nbreaks <- length(x.f$breaks) x.bin <- x.f$breaks[-1]-x.f$breaks[-nbreaks] x.mid <- (x.f$breaks[-nbreaks]+x.f$breaks[-1])/2 x.mean <- sum(x.mid*x.f$density*x.bin) x.sd <- sqrt(sum(x.mid*x.mid*x.f$density*x.bin)-x.mean^2) return(list(mean=x.mean, sd=x.sd)) } get.hist.var <- function(x.f){ nbreaks <- length(x.f$breaks) x.bin <- x.f$breaks[-1]-x.f$breaks[-nbreaks] x.mid <- (x.f$breaks[-nbreaks]+x.f$breaks[-1])/2 x.mean <- sum(x.mid*x.f$density*x.bin) return(mean=x.mean) } # obsolete function (01-06-2010) # plot plot.ez.group.old <- function(ez.list, plot.dir, file.name=NULL, legend.txt, y.lim=NULL, xlab.txt="num of significant peaks", ylab.txt="avg posterior prob of being random", col.txt=NULL, title.txt=NULL){ if(is.null(col.txt)) col.txt <- c("black", "red", "purple", "green", "blue", "cyan", "magenta", "orange", "grey") x <- c() y <- c() for(i in 1:length(ez.list)){ x <- c(x, ez.list[[i]]$n) y <- c(y, ez.list[[i]]$uri.ez) } if(is.null(y.lim)) y.lim <- c(0, max(y)) if(!is.null(file.name)){ postscript(paste(plot.dir, "ez.", file.name, sep="")) par(mfrow=c(1,1), mar=c(5,5,4,2)) } plot(x, y, ylim=y.lim, type="n", xlab=xlab.txt, ylab=ylab.txt, lwd=5, cex=5, cex.axis=2, cex.lab=2) for(i in 1:length(ez.list)){ lines(ez.list[[i]]$n, ez.list[[i]]$uri.ez, col=col.txt[i], cex=2, lwd=5) } # plot(ez.list[[1]]$u.t, y, ylim=y.lim, type="l", xlab="rep-sig", ylab=ylab.txt, lwd=5, cex=5, cex.axis=2, cex.lab=2) # plot(ez.list[[1]]$v.t, y, ylim=y.lim, type="l", xlab="rep-sig", ylab=ylab.txt, lwd=5, cex=5, cex.axis=2, cex.lab=2) legend(0, y.lim[2], legend=legend.txt, col=col.txt[1:length(col.txt)], lty=1, lwd=5, cex=2) if(!is.null(title)) title(title.txt) if(!is.null(file.name)){ dev.off() } } plot.ez.group <- function(ez.list, plot.dir, file.name=NULL, legend.txt, y.lim=NULL, xlab.txt="num of significant peaks", ylab.txt="IDR", col.txt=NULL, title.txt=NULL){ if(is.null(col.txt)) col.txt <- c("black", "red", "purple", "green", "blue", "cyan", "magenta", "orange", "grey") n.entry <- length(ez.list) x <- rep(NA, n.entry) y.max <- rep(NA, n.entry) for(i in 1:n.entry){ x[i] <- max(ez.list[[i]]$n) y.max[i] <- max(ez.list[[i]]$IDR) } if(is.null(y.lim)) y.lim <- c(0, max(y.max)) if(!is.null(file.name)){ postscript(paste(plot.dir, "ez.", file.name, sep="")) par(mfrow=c(1,1), mar=c(5,5,4,2)) } plot(c(0, max(x)), y.lim, ylim=y.lim, type="n", xlab=xlab.txt, ylab=ylab.txt, lwd=5, cex=5, cex.axis=2, cex.lab=2) q <- seq(0.01, 0.99, by=0.01) for(i in 1:length(ez.list)){ n.plot <- round(quantile(ez.list[[i]]$n, prob=q)) IDR.plot <- ez.list[[i]]$IDR[n.plot] lines(n.plot, IDR.plot, col=col.txt[i], cex=2, lwd=5) } legend(0, y.lim[2], legend=legend.txt, col=col.txt[1:length(col.txt)], lty=1, lwd=5, cex=2) if(!is.null(title)) title(title.txt) if(!is.null(file.name)){ dev.off() } } ############################################################################# ############################################################################# # statistics about peaks selected on the individual replicates # # idr.level: the consistency cutoff, say 0.05 # uri.output: a list of uri.output from consistency analysis generated by batch-consistency-analysis.r # ez.list : a list of IDRs computed from get.ez.tt using the same idr.level # ################## # obsolete? # compute the error rate # u.t and v.t are the quantiles # # map back to all peaks and report the number of peaks selected get.ez.tt.all.old <- function(em.fit, all.data1, all.data2, idr.level){ u <- em.fit$data.pruned$sample1 v <- em.fit$data.pruned$sample2 tt <- seq(0.01, 0.99, by=0.01) # if(reverse){ e.z <- 1-em.fit$em.fit$e.z # this is the error prob uri.ez <- sapply(tt, comp.uri.ez, u=u, v=v, e.z=e.z) ez.bound <- sapply(tt, comp.ez.cutoff, u=u, v=v, e.z=e.z, boundary.txt="max") # } else { # e.z <- em.fit$em.fit$e.z # uri.ez <- sapply(tt, comp.uri.ez, u=u, v=v, e.z=e.z) # ez.bound <- sapply(tt, comp.ez.cutoff, u=u, v=v, e.z=e.z, boundary.txt="min") # } u.t <- quantile(u, prob=(1-tt)) v.t <- quantile(v, prob=(1-tt)) # find the levels on the two replicates sig.value1 <- c() sig.value2 <- c() error.prob.cutoff <- c() n.selected.match <- c() npeak.rep1 <- c() npeak.rep2 <- c() for(i in 1:length(idr.level)){ # find which uri.ez is closet to idr.level index <- which.min(abs(uri.ez - as.numeric(idr.level[i]))) sig.value1[i] <- u.t[index] sig.value2[i] <- v.t[index] error.prob.cutoff[i] <- ez.bound[index] n.selected.match[i] <- sum(u>= u.t[index] & v>=v.t[index]) npeak.rep1[i] <- sum(all.data1["sig.value"] >= sig.value1[i]) npeak.rep2[i] <- sum(all.data2["sig.value"] >= sig.value2[i]) } # output the cutoff of posterior probability, signal values on two replicates map.uv <- cbind(error.prob.cutoff, sig.value1, sig.value2, n.selected.match, npeak.rep1, npeak.rep2) return(list(n=tt*length(u), uri.ez=uri.ez, u.t=u.t, v.t=v.t, tt=tt, idr.level=idr.level, map.uv=map.uv, e.z=e.z, error.prob.cutoff=error.prob.cutoff)) } get.ez.tt.all <- function(em.fit, all.data1, all.data2, idr.level=c(0.01, 0.05, 0.1)){ u <- em.fit$data.pruned$sample1$sig.value v <- em.fit$data.pruned$sample2$sig.value # u <- em.fit$data.pruned$sample1 # v <- em.fit$data.pruned$sample2 e.z <- 1-em.fit$em.fit$e.z # this is the error prob o <- order(e.z) e.z.ordered <- e.z[o] n.select <- c(1:length(e.z)) IDR <- cumsum(e.z.ordered)/n.select u.o <- u[o] v.o <- v[o] n.level <- length(idr.level) # sig.value1 <- rep(NA, n.level) # sig.value2 <- rep(NA, n.level) ez.cutoff <- rep(NA, n.level) n.selected <- rep(NA, n.level) npeak.rep1 <- rep(NA, n.level) npeak.rep2 <- rep(NA, n.level) for(i in 1:length(idr.level)){ # find which uri.ez is closet to fdr.level index <- which.min(abs(IDR - idr.level[i])) # sig.value1[i] <- min(u.o[1:index]) # sig.value2[i] <- min(v.o[1:index]) ez.cutoff[i] <- e.z.ordered[index] # fixed on 02/20/10 n.selected[i] <- sum(e.z<=ez.cutoff[i]) # npeak.rep1[i] <- sum(all.data1["sig.value"] >= sig.value1[i]) # npeak.rep2[i] <- sum(all.data2["sig.value"] >= sig.value2[i]) } # output the cutoff of posterior probability, number of selected overlapped peaks map.uv <- cbind(ez.cutoff, n.selected) return(list(n=n.select, IDR=IDR, idr.level=idr.level, map.uv=map.uv)) } # return(list(n=tt*length(u), uri.ez=uri.ez, fdr.level=fdr.level, map.uv=map.uv, e.z=e.z, error.prob.cutoff=error.prob.cutoff)) ####### the following is for determining thresholds for merged dataset ############# select peaks above a given threshold # # pass.threshold: a simple method, passing the threshold on the threshold on the individual replicate to the pooled sample # # sig.map.list: a list of matrix to include all the cutoff values, each row corresponds to a cutoff. The first column is idr.level # the 2nd column is the cutoff of ez, the rest of columns are consistency analysis for other replicates # sig.value.name: the name of the sig.value column # combined: combined dataset # nrep: number of pairs of comparisons # # Procedure: # 1. Find the significant threshold corresponding to the idr cutoff on the matched peaks. # 2. Each time we will get two or more (if >2 replicates) cutoffs and will report the most stringent and the least stringent # cutoff and the number of peaks selected at those two cutoffs ############# pass.threshold <- function(sig.map.list, sig.value.name, combined, idr.level, nrep, chr.size){ sig.map <- c() # choose idr.level idr.index <- which(rbind(sig.map.list[[1]])[,1] == idr.level) if(length(i) ==0){ print("no level matches specified idr.level") return(-1) } for(i in 1:length(sig.map.list)) sig.map <- c(sig.map, rbind(sig.map.list[[i]])[idr.index, c("sig.value1", "sig.value2")]) npeak.tight <- c() npeak.loose <- c() max.sig <- max(sig.map) min.sig <- min(sig.map) selected.sig.tight <- combined[combined[,sig.value.name]>=max.sig, ] selected.sig.loose <- combined[combined[,sig.value.name]>=min.sig, ] selected.sig.tight <- deconcatenate.chr(selected.sig.tight, chr.size)[,c("chr", "start", "stop", "signal.value", "p.value", "q.value")] selected.sig.loose <- deconcatenate.chr(selected.sig.loose, chr.size)[,c("chr", "start", "stop", "signal.value", "p.value", "q.value")] npeak.tight <- nrow(selected.sig.tight) npeak.loose <- nrow(selected.sig.loose) npeak.stat <- list(idr.level=idr.level, max.sig=max.sig, min.sig=min.sig, npeak.tight=npeak.tight, npeak.loose=npeak.loose) invisible(list(npeak.stat=npeak.stat, combined.selected.tight=selected.sig.tight, combined.selected.loose=selected.sig.loose)) } ################# # pass the regions selected from consistency analysis to combined data # Threshold is determined on the replicates, the regions above the threshold are selected # then peaks on the combined data are selected from the selected regions # # To avoid being too stringent, regions satisfying the following conditions are selected # 1. regions above the significant threshold determined by consistency analysis on either replicate # 2. regions that have consistent low peaks, i.e. posterior prob > threshold but not passing the significant threshold # # This method doesn't make a difference when using different thresholds ################# pass.region <- function(sig.map.list, uri.output, ez.list, em.output, combined, idr.level, sig.value.impute=0, chr.size){ combined <- combined[, c("start", "stop", "sig.value", "signal.value", "p.value", "q.value")] npair <- length(uri.output) # number of pairs of consistency analysis combined.region <- c() # choose idr.level idr.index <- which(rbind(sig.map.list[[1]])[,1] == idr.level) if(length(idr.index) ==0){ print("no level matches specified idr.level") return(-1) } for(j in 1:npair){ # select peaks from individual replicates using individual cutoff above.1 <- uri.output[[j]]$data12.enrich$merge1["sig.value"] >= ez.list[[j]]$map.uv[idr.index,"sig.value1"] above.2 <- uri.output[[j]]$data12.enrich$merge1["sig.value"] >= ez.list[[j]]$map.uv[idr.index,"sig.value2"] selected.sig.rep1 <- uri.output[[j]]$data12.enrich$merge1[above.1, c("start", "stop", "sig.value", "signal.value", "p.value", "q.value")] selected.sig.rep2 <- uri.output[[j]]$data12.enrich$merge2[above.2, c("start", "stop", "sig.value", "signal.value", "p.value", "q.value")] # find the peaks that are overlapped with reliable peaks in the individual replicates overlap.1 <- pair.peaks(selected.sig.rep1, combined)$merge2 overlap.2 <- pair.peaks(selected.sig.rep2, combined)$merge2 # choose the ones with significant value > 0, which are the overlapped ones combined.in1 <- overlap.1[overlap.1$sig.value > sig.value.impute, c("start", "stop", "sig.value", "signal.value", "p.value", "q.value")] combined.in2 <- overlap.2[overlap.2$sig.value > sig.value.impute, c("start", "stop", "sig.value", "signal.value", "p.value", "q.value")] ## consistent low significant ones ## first find consistenct ones, ie. high posterior prob # is.consistent <- ez.list[[j]]$e.z < ez.list[[j]]$ez.cutoff # data.matched <- keep.match(uri.output[[j]]$data12.enrich$merge1[!above.1, ], uri.output[[j]]$data12.enrich$merge2[!above.2, ], sig.value.impute=0) # data.matched$sample1 <- data.matched$sample1[, c("start", "stop", "sig.value", "signal.value", "p.value", "q.value")] # data.matched$sample2 <- data.matched$sample2[, c("start", "stop", "sig.value", "signal.value", "p.value", "q.value")] # consistent.in1 <- data.matched$sample1[is.consistent, ] # consistent.in2 <- data.matched$sample2[is.consistent, ] # overlap.consistent.1 <- pair.peaks(consistent.in1, combined)$merge2 # overlap.consistent.2 <- pair.peaks(consistent.in2, combined)$merge2 ## choose the ones with significant value > 0, which are the overlapped ones # combined.consistent.in1 <- overlap.consistent.1[overlap.consistent.1$sig.value > sig.value.impute, ] # combined.consistent.in2 <- overlap.consistent.2[overlap.consistent.2$sig.value > sig.value.impute, ] # combined.region <- rbind(combined.region, combined.in1, combined.in2, combined.consistent.in1, combined.consistent.in2) combined.region <- rbind(combined.region, combined.in1, combined.in2) is.repeated <- duplicated(combined.region$start) combined.region <- combined.region[!is.repeated, c("start", "stop", "sig.value", "signal.value", "p.value", "q.value")] } npeak <- nrow(combined.region) sig.combined <- c(min(combined.region[,"sig.value"], na.rm=T), max(combined.region[,"sig.value"], na.rm=T)) # idr.combined <- c(min(combined.region[,"q.value"], na.rm=T), max(combined.region[,"q.value"], na.rm=T)) npeak.stat <- list(idr.level=idr.level, npeak=npeak) combined.region <- deconcatenate.chr(combined.region, chr.size)[,c("chr", "start", "stop", "signal.value", "p.value", "q.value")] invisible(list(npeak.stat=npeak.stat, combined.selected=combined.region, sig.combined=sig.combined)) } ################ # pass structure: this method does another round of inference on the combined data # # To make the mixture structure comparable on the replicates and the combined data, the 2nd inference is done on the peaks # at the reliable regions on the combined data, using rank transformed significant values. The mixture structure is estimated using my consistency analysis, which # estimates marginal distributions of ranks using nonparametric ways. Then the significant values are found out. # There are several advantages to do it this way: # 1. The premise of passing structure is that the means and variance (i.e. distribution) of two replicates should be the same # The significant values on the two replicates clearly have different distributions. The structure estimated from consistency # analysis will generate similar rank distribution on two replicates by its setup (i.e. same number of peaks are paired up). # 2. Because pooled sample is a black box, the structure is more likely to be followed in the matched regions than other locations, # after all, we don't know what other things are. If even the structure doesn't hold on the matched regions, # which is possible, let alone the other regions. Focusing on the reliable regions helps to get rid of those unknown noises. # # # modified on 2-20-10: reverse rank.combined, make big sig.value with small # ranks, to be consistent with f1 and f2 ################ pass.structure <- function(uri.output, em.output, combined, idr.level, sig.value.impute, chr.size, overlap.ratio=0){ columns.keep <- c("sig.value", "start", "stop", "signal.value", "p.value", "q.value", "chr", "start.ori", "stop.ori") combined <- combined[, columns.keep] combined.selected.all <- c() for(j in 1:npair){ sample1 <- uri.output[[j]]$data12.enrich$merge1[, columns.keep] sample2 <- uri.output[[j]]$data12.enrich$merge2[, columns.keep] # find peaks on the matched region on the combined one data.matched <- keep.match(sample1, sample2, sig.value.impute=sig.value.impute) data.matched$sample1 <- data.matched$sample1[, columns.keep] data.matched$sample2 <- data.matched$sample2[, columns.keep] overlap.1 <- pair.peaks.filter(data.matched$sample1, combined, p.value.impute=sig.value.impute, overlap.ratio)$merge2 overlap.2 <- pair.peaks.filter(data.matched$sample2, combined, p.value.impute=sig.value.impute, overlap.ratio)$merge2 # choose the ones with significant value > sig.value.impute, which are the overlapped ones combined.in1 <- overlap.1[overlap.1$sig.value > sig.value.impute, ] combined.in2 <- overlap.2[overlap.2$sig.value > sig.value.impute, ] combined.region <- rbind(combined.in1, combined.in2) is.repeated <- duplicated(combined.region$start) combined.region <- combined.region[!is.repeated,] # now rank the peaks in matched region rank.combined <- rank(-combined.region$sig.value) # now transform the parameters estimated into the new scale npeaks.overlap <- nrow(combined.region) npeaks.consistent <- nrow(cbind(em.output[[j]]$data.pruned$sample1)) # the breaks are the same for x and y f1 <- list(breaks=em.output[[j]]$em.fit$x.mar$f1$breaks*npeaks.overlap/npeaks.consistent, density=(em.output[[j]]$em.fit$x.mar$f1$density+em.output[[j]]$em.fit$y.mar$f1$density)/2) # the first break boundary goes up when changing scale, need set it back to be a bit smaller than 1 f1$breaks[1] <- min(f1$breaks[1], 0.95) f2 <- list(breaks=em.output[[j]]$em.fit$x.mar$f2$breaks*npeaks.overlap/npeaks.consistent, density=(em.output[[j]]$em.fit$x.mar$f2$density+em.output[[j]]$em.fit$y.mar$f2$density)/2) # the first break boundary goes up when changing scale, need set it back to be a bit smaller than 1 f2$breaks[1] <- min(f2$breaks[1], 0.95) p <- em.output[[j]]$em.fit$para$p # find the posterior probability errorprob.combined <- get.comp2.prob(rank.combined, p, f1, f2) # compute the FDR and find cutoff of posterior prob and the sig value o <- order(errorprob.combined) idr <- cumsum(errorprob.combined[o])/c(1:length(o)) idr.index <- which(idr > idr.level)[1] errorprob.cutoff <- errorprob.combined[o][idr.index] # find the minimum significant measure among selected peaks sig.value <- min(combined.region$sig.value[o][1:idr.index]) # sig.value <- quantile(combined.region$sig.value[o][1:idr.index], prob=0.05) #sig.value <- quantile(combined.region$sig.value[errorprob.combined<=errorprob.cutoff], prob=0.05) # apply the significant value on the whole pooled list combined.selected <- combined[combined$sig.value >= sig.value,] combined.selected.all <- rbind(combined.selected.all, combined.selected) } is.repeated <- duplicated(combined.selected.all$start) combined.selected.all <- combined.selected.all[!is.repeated,] npeak <- nrow(combined.selected.all) npeak.stat <- list(idr.level=idr.level, npeak=npeak) sig.combined <- c(min(combined.selected.all[,"sig.value"], na.rm=T), max(combined.selected.all[,"sig.value"], na.rm=T)) # idr.combined <- c(min(combined.selected.all[,"q.value"], na.rm=T), max(combined.selected.all[,"q.value"], na.rm=T)) # combined.selected.all <- deconcatenate.chr(combined.selected.all, chr.size)[,c("chr", "start", "stop", "signal.value", "p.value", "q.value")] combined.selected.all <- combined.selected.all[, c("chr", "start.ori", "stop.ori", "signal.value", "p.value", "q.value")] colnames(combined.selected.all) <- c("chr", "start", "stop", "signal.value", "p.value", "q.value") invisible(list(npeak.stat=npeak.stat, combined.selected=combined.selected.all, sig.combined=sig.combined)) } # get the posterior probability of the 2nd component get.comp2.prob <- function(x, p, f1, f2){ # get pdf and cdf of each component from functions in the corresponding component px.1 <- sapply(x, get.pdf, df=f1) px.2 <- sapply(x, get.pdf, df=f2) comp2prob <- 1 - p*px.1/(p*px.1+(1-p)*px.2) return(comp2prob) } keep.match <- function(sample1, sample2, sig.value.impute=0){ sample1.prune <- sample1[sample1$sig.value > sig.value.impute & sample2$sig.value > sig.value.impute,] sample2.prune <- sample2[sample1$sig.value > sig.value.impute & sample2$sig.value > sig.value.impute,] invisible(list(sample1=sample1.prune, sample2=sample2.prune)) } ############################################## # # The following is for simulation # ############################################## # simulate gaussian copula # u is the uniform random variable and rho is correlation coefficient simu.gaussian.copula <- function(u, rho){ n <- length(u) # simulate y given x=qnorm(u) y <- qnorm(u)*rho + rnorm(n)*sqrt(1-rho^2) v <- pnorm(y) invisible(v) } ## simulate Clayton copula from its generating function ## Genest and MacKay (1986) phi.ori <- function(t, s){ (t^(-s) -1)/s } phi.inv <- function(y, s){ exp(-log(s*y+1)/s) } phi.der <- function(t, s){ -t^(-s-1) } phi.der.inv <- function(y, s){ exp(log(-y)/(-s-1)) } get.w <- function(u, t, s){ phi.der.inv(phi.der(u, s)/t, s) } get.v <- function(w, u, s){ phi.inv(phi.ori(w, s) - phi.ori(u, s), s) } # u is a uniform random variable, s is the association parameter simu.clayton.copula <- function(u, s){ t <- runif(length(u)) if(s>0){ w <- get.w(u, t, s) v <- get.v(w, u, s) return(v) } if(s==0){ return(t) } if(s <0){ print("Invalid association parameters for clayton copula") } } ###### 09-09-09 # simulate a two-component copula mixture: # - marginal distributions for the two variables in each component are both # normal and with the same parameters # p is the mixing proportion of component 1 # n is the total sample size simu.copula.2mix <- function(s1, s2, p, n, mu1, mu2, sd1, sd2, copula.txt){ n1 <- round(n*p) n2 <- n-n1 u1 <- runif(n1) if(copula.txt =="clayton") v1 <- simu.clayton.copula(u1, s1) else{ if(copula.txt =="gaussian") v1 <- simu.gaussian.copula(u1, s1) } u2 <- runif(n2) if(copula.txt =="clayton") v2 <- simu.clayton.copula(u2, s2) else{ if(copula.txt =="gaussian") v2 <- simu.gaussian.copula(u2, s2) } # generate test statistics sample1.1 <- qnorm(u1, mu1, sd1) sample1.2 <- qnorm(v1, mu1, sd1) sample2.1 <- qnorm(u2, mu2, sd2) sample2.2 <- qnorm(v2, mu2, sd2) return(list(u=c(u1, u2), v=c(v1, v2), u.inv=c(sample1.1, sample2.1), v.inv=c(sample1.2, sample2.2), label=c(rep(1, n1), rep(2, n2)))) } # using inverse of the cdf to generate original observations simu.copula.2mix.inv <- function(s1, s2, p, n, cdf1.x, cdf1.y, cdf2.x, cdf2.y, copula.txt){ n1 <- round(n*p) n2 <- n-n1 u1 <- runif(n1) if(copula.txt =="clayton") v1 <- simu.clayton.copula(u1, s1) else{ if(copula.txt =="gaussian") v1 <- simu.gaussian.copula(u1, s1) } u2 <- runif(n2) if(copula.txt =="clayton") v2 <- simu.clayton.copula(u2, s2) else{ if(copula.txt =="gaussian") v2 <- simu.gaussian.copula(u2, s2) } # generate test statistics # sample1.1 <- qnorm(u1, mu1, sd1) # sample1.2 <- qnorm(v1, mu1, sd1) # sample2.1 <- qnorm(u2, mu2, sd2) # sample2.2 <- qnorm(v2, mu2, sd2) sample1.x <- inv.cdf.vec(u1, cdf1.x) sample1.y <- inv.cdf.vec(v1, cdf1.y) sample2.x <- inv.cdf.vec(u2, cdf2.x) sample2.y <- inv.cdf.vec(v2, cdf2.y) return(list(u=c(u1, u2), v=c(v1, v2), u.inv=c(sample1.x, sample2.x), v.inv=c(sample1.y, sample2.y), label=c(rep(1, n1), rep(2, n2)))) } # obtain original observation by converting cdf into quantiles # u is one cdf # u.cdf is a cdf (assuming it is a histogram) and has the break points (cdf$cdf and cdf$breaks) # the smallest value of cdf=0 and the largest =1 inv.cdf <- function(u, u.cdf){ # which bin it falls into i <- which(u.cdf$cdf> u)[1] q.u <- (u - u.cdf$cdf[i-1])/(u.cdf$cdf[i] - u.cdf$cdf[i-1])* (u.cdf$breaks[i]-u.cdf$breaks[i-1]) + u.cdf$breaks[i-1] return(q.u) } inv.cdf.vec <- function(u, u.cdf){ # check if cdf has the right range (0, 1) ncdf <- length(u.cdf$cdf) nbreaks <- length(u.cdf$breaks) if(ncdf == nbreaks-1 & u.cdf$cdf[ncdf]< 1) u.cdf[ncdf] <- 1 q.u <- sapply(u, inv.cdf, u.cdf) return(q.u) } # here we simulate a likely real situation # the test statistics from two normal distributions # according to their labels, then convert them into p-values w.r.t H0 using # one-sided test. # The test statistics are correlated for the signal component and independent # for the noise component # For the signal component, Y = X + eps, where eps ~ N(0, sigma^2) simu.test.stat <- function(p, n, mu1, sd1, mu0, sd0, sd.e){ # first component - signal n.signal <- round(n*p) n.noise <- n - n.signal # labels labels <- c(rep(1, n.signal), rep(0, n.noise)) # test statistics for signal and noise mu.signal <- rnorm(n.signal, mu1, sd1) x.signal <- mu.signal + rnorm(n.signal, 0, sd.e) x.noise <- rnorm(n.noise, mu0, sd0) + rnorm(n.noise, 0, sd.e) y.signal <- mu.signal + rnorm(n.signal, 0, sd.e) # sd.e can be dependent on signal y.noise <- rnorm(n.noise, mu0, sd0) + rnorm(n.noise, 0, sd.e) # concatenate x <- c(x.signal, x.noise) y <- c(y.signal, y.noise) # convert to p-values based on H0 p.x <- 1-pnorm(x, mu0, sqrt(sd0^2+sd.e^2)) p.y <- 1-pnorm(y, mu0, sqrt(sd0^2+sd.e^2)) return(list(p.x=p.x, p.y=p.y, x=x, y=y, labels=labels)) } # compute the tradeoff and calibration forward.decoy.tradeoff.ndecoy <- function(xx, labels, ndecoy){ xx <- round(xx, 5) o <- order(xx, decreasing=T) rand <- 1-labels # if rand==0, consistent # order the random indicator in the same order rand.o <- rand[o] if(sum(rand.o) > ndecoy){ index.decoy <- which(cumsum(rand.o)==ndecoy) } else { index.decoy <- which(cumsum(rand.o)==sum(rand.o)) } cutoff.decoy <- xx[o][index.decoy] # only consider the unique ones cutoff.unique <- unique(xx[o]) cutoff <- cutoff.unique[cutoff.unique >= cutoff.decoy[length(cutoff.decoy)]] get.decoy.count <- function(cut.off){ above <- rep(0, length(xx)) above[xx >= cut.off] <- 1 decoy.count <- sum(above==1 & rand==1) return(decoy.count) } get.forward.count <- function(cut.off){ above <- rep(0, length(xx)) above[xx >= cut.off] <- 1 forward.count <- sum(above==1 & rand==0) return(forward.count) } get.est.fdr <- function(cut.off){ above <- rep(0, length(xx)) above[xx >= cut.off] <- 1 est.fdr <- 1-mean(xx[above==1]) return(est.fdr) } # assuming rand=0 is right get.false.neg.count <- function(cut.off){ below <- rep(0, length(xx)) below[xx < cut.off] <- 1 false.neg.count <- sum(below==1 & rand==0) return(false.neg.count) } get.false.pos.count <- function(cut.off){ above <- rep(0, length(xx)) above[xx >= cut.off] <- 1 false.pos.count <- sum(above==1 & rand==1) return(false.pos.count) } decoy <- sapply(cutoff, get.decoy.count) forward <- sapply(cutoff, get.forward.count) est.fdr <- sapply(cutoff, get.est.fdr) emp.fdr <- decoy/(decoy+forward) # compute specificity and sensitivity # assuming rand=1 is wrong and rand=0 is right false.neg <- sapply(cutoff, get.false.neg.count) false.pos <- sapply(cutoff, get.false.pos.count) true.pos <- sum(rand==0)-false.neg true.neg <- sum(rand==1)-false.pos sensitivity <- true.pos/(true.pos+false.neg) specificity <- true.neg/(true.neg+false.pos) return(list(decoy=decoy, forward=forward, cutoff=cutoff, est.fdr=est.fdr, emp.fdr=emp.fdr, sensitivity=sensitivity, specificity=specificity)) } # compute the em for jackknife and all data, and find FDR get.emp.jack <- function(a, p0){ nobs <- length(a$labels) est <- list() est.all <- list() temp.all <- em.transform(-a$p.x, -a$p.y, mu=1.5, sigma=1.4, rho=0.4, p=0.7, eps=0.01) # temp.all <- em.2copula.quick(a$p.x, a$p.y, p0=p0, rho1.0=0.7, # rho2.0=0, eps=0.01, fix.p=T, stoc=F, fix.rho2=T, "gaussian") est.all$p <- temp.all$para$p est.all$rho1 <- temp.all$para$rho1 est.all$FDR <- get.FDR(temp.all$e.z) FDR <- list() p <- c() rho1 <- c() for(i in 1:nobs){ temp <- em.transform(-a$p.x[-i], -a$p.y[-i], mu=1.5, sigma=1.4, rho=0.4, p=0.7, eps=0.01) # temp <- em.2copula.quick(a$p.x[-i], a$p.y[-i], p0=p0, rho1.0=0.7, # rho2.0=0, eps=0.01, fix.p=T, stoc=F, fix.rho2=T, "gaussian") est[[i]] <- list(p=temp$para$p, rho1=temp$para$rho1, FDR=get.FDR(temp$e.z)) FDR[[i]] <- est[[i]]$FDR # this is the FDR for top n peaks p[i] <- est[[i]]$p rho1[i] <- est[[i]]$rho1 } est.jack <- list(FDR=FDR, p=p, rho1=rho1) return(list(est.jack=est.jack, est.all=est.all)) } # get the npeaks corresponding to the nominal FDR estimated from the sample # and find the corresponding FDR from the entire data get.FDR.jack <- function(est, FDR.nominal){ nobs <- length(est$est.jack$FDR) FDR.all <- c() top.n <- c() for(i in 1:nobs){ top.n[i] <- max(which(est$est.jack$FDR[[i]] <= FDR.nominal)) FDR.all[i] <- est$est.all$FDR[top.n[i]] } invisible(list(FDR.all=FDR.all, top.n=top.n)) } # compute Jackknife peudonumber # a is the dataset get.emp.IF <- function(a, p0){ nobs <- length(a$labels) est <- list() est.all <- list() temp.all <- em.2copula.quick(a$p.x, a$p.y, p0=p0, rho1.0=0.7, rho2.0=0, eps=0.01, fix.p=T, stoc=F, fix.rho2=T, "gaussian") est.all$p <- temp.all$para$p est.all$rho1 <- temp.all$para$rho1 est.all$FDR <- get.FDR(temp.all$e.z) IF.FDR <- list() IF.p <- c() IF.rho1 <- c() for(i in 1:nobs){ temp <- em.2copula.quick(a$p.x[-i], a$p.y[-i], p0=p0, rho1.0=0.7, rho2.0=0, eps=0.01, fix.p=T, stoc=F, fix.rho2=T, "gaussian") est[[i]] <- list(p=temp$para$p, rho1=temp$para$rho1, FDR=get.FDR(temp$e.z)) IF.FDR[[i]] <- (nobs-1)*(est.all$FDR[-nobs] - est[[i]]$FDR) # this is the FDR for top n peaks IF.p[i] <- (nobs-1)*(est.all$p - est[[i]]$p) IF.rho1[i] <- (nobs-1)*(est.all$rho1 - est[[i]]$rho1) } emp.IF <- list(FDR=IF.FDR, p=IF.p, rho1=IF.rho1) invisible(list(emp.IF=emp.IF, est.all=est.all, est=est)) } # e.z is the posterior probability of being in signal component get.FDR <- function(e.z){ e.z.o <- order(1-e.z) FDR <- cumsum(1-e.z[e.z.o])/c(1:length(e.z.o)) invisible(FDR) } # get the FDR of selecting the top n peaks # IF.est is the sample influence function # top.n get.IF.FDR <- function(IF.est, top.n){ nobs <- length(IF.est$emp.IF$FDR) FDR <- c() # influence function of p for(i in 1:nobs) FDR[i] <- IF.est$emp.IF$FDR[[i]][top.n] invisible(FDR) } # get the sample influence function for FDR at a given FDR size # 1. find the number of peaks selected at a given FDR computed from all obs # 2. use the number to find the sample influence function for FDR # IF.est$est.all is the FDR with all peaks get.IF.FDR.all <- function(IF.est, FDR.size){ top.n <- which.min(abs(IF.est$est.all$FDR -FDR.size)) nobs <- length(IF.est$est.all$FDR) FDR <- c() # influence function of p for(i in 1:nobs) FDR[i] <- IF.est$emp.IF$FDR[[i]][top.n] invisible(list(FDR=FDR, top.n=top.n)) } plot.simu.uri <- function(x, y){ tt <- seq(0.01, 0.99, by=0.01) uri <- sapply(tt, comp.uri.prob, u=x, v=y) uri.thin <- uri[seq(1, length(tt), by=3)] tt.thin <- tt[seq(1, length(tt), by=3)] duri <- (uri.thin[-1]-uri.thin[-length(uri.thin)])/(tt.thin[-1]-tt.thin[-length(tt.thin)]) uri.spl <- smooth.spline(tt, uri, df=6.4) uri.der <- predict(uri.spl, tt, deriv=1) par(mfrow=c(2,2)) plot(x[1:n0], y[1:n0]) points(x[(n0+1):n], y[(n0+1):n], col=2) plot(rank(-x)[1:n0], rank(-y)[1:n0]) points(rank(-x)[(1+n0):n], rank(-y)[(1+n0):n]) plot(tt, uri) lines(c(0,1), c(0,1), lty=2) title(paste("rho1=", rho1, " rho2=", rho2, "p=", p, sep="")) plot(tt.thin[-1], duri) lines(uri.der) abline(h=1) invisible(list(x=x, y=y, uri=uri, tt=tt, duri=duri, tt.thin=tt.thin, uri.der=uri.der)) } ###### new fitting procedure # 1. rank pairs # 2. initialization # 3. convert to pseudo-number # 4. EM # need plugin and test # find the middle point between the bins get.pseudo.mix <- function(x, mu, sigma, rho, p){ # first compute cdf for points on the grid # generate 200 points between [-3, mu+3*sigma] nw <- 1000 w <- seq(min(-3, mu-3*sigma), max(mu+3*sigma, 3), length=nw) w.cdf <- p*pnorm(w, mean=mu, sd=sigma) + (1-p)*pnorm(w, mean=0, sd=1) i <- 1 quan.x <- rep(NA, length(x)) for(i in c(1:nw)){ index <- which(x >= w.cdf[i] & x < w.cdf[i+1]) quan.x[index] <- (x[index]-w.cdf[i])*(w[i+1]-w[i])/(w.cdf[i+1]-w.cdf[i]) +w[i] } index <- which(x < w.cdf[1]) if(length(index)>0) quan.x[index] <- w[1] index <- which(x > w.cdf[nw]) if(length(index)>0) quan.x[index] <- w[nw] # linear.ext <- function(x, w, w.cdf){ # linear interpolation # index.up <- which(w.cdf>= x)[1] # left.index <- which(w.cdf <=x) # index.down <- left.index[length(left.index)] # quan.x <- (w[index.up] + w[index.down])/2 # } # x.pseudo <- sapply(x, linear.ext, w=w, w.cdf=w.cdf) # invisible(x.pseudo) invisible(quan.x) } # EM to compute the latent structure # steps: # 1. raw values are first transformed into pseudovalues # 2. EM is used to compute the underlining structure, which is a mixture # of two normals em.transform <- function(x, y, mu, sigma, rho, p, eps){ x.cdf.func <- ecdf(x) y.cdf.func <- ecdf(y) afactor <- length(x)/(length(x)+1) x.cdf <- x.cdf.func(x)*afactor y.cdf <- y.cdf.func(y)*afactor # initialization para <- list() para$mu <- mu para$sigma <- sigma para$rho <- rho para$p <- p j <- 1 to.run <- T loglik.trace <- c() loglik.inner.trace <- c() #to.run.inner <- T z.1 <- get.pseudo.mix(x.cdf, para$mu, para$sigma, para$rho, para$p) z.2 <- get.pseudo.mix(y.cdf, para$mu, para$sigma, para$rho, para$p) # cat("length(z1)", length(z.1), "\n") while(to.run){ # get pseudo value in each cycle # z.1 <- get.pseudo.mix(x.cdf, para$mu, para$sigma, para$rho, para$p) # z.2 <- get.pseudo.mix(y.cdf, para$mu, para$sigma, para$rho, para$p) i <- 1 while(to.run){ # EM for latent structure e.z <- e.step.2normal(z.1, z.2, para$mu, para$sigma, para$rho, para$p) para <- m.step.2normal(z.1, z.2, e.z) #para$rho <- rho #para$p <- p #para$mu <- mu #para$sigma <- sigma if(i > 1) l.old <- l.new # this is just the mixture likelihood of two-component Gaussian l.new <- loglik.2binormal(z.1, z.2, para$mu, para$sigma, para$rho, para$p) loglik.inner.trace[i] <- l.new if(i > 1){ to.run <- loglik.inner.trace[i]-loglik.inner.trace[i-1]>eps } # if(i > 2){ # l.inf <- loglik.inner.trace[i-2] + (loglik.inner.trace[i-1] - loglik.inner.trace[i-2])/(1-(loglik.inner.trace[i]-loglik.inner.trace[i-1])/(loglik.inner.trace[i-1]-loglik.inner.trace[i-2])) # if(loglik.inner.trace[i-1]!=loglik.inner.trace[i-2]) # to.run <- abs(l.inf - loglik.inner.trace[i]) > eps # else # to.run <- F # } cat("loglik.inner.trace[", i, "]=", loglik.inner.trace[i], "\n") cat("mu=", para$mu, "sigma=", para$sigma, "p=", para$p, "rho=", para$rho, "\n\n") i <- i+1 } # get pseudo value in each cycle z.1 <- get.pseudo.mix(x.cdf, para$mu, para$sigma, para$rho, para$p) z.2 <- get.pseudo.mix(y.cdf, para$mu, para$sigma, para$rho, para$p) if(j > 1) l.old.outer <- l.new.outer l.new.outer <- loglik.2binormal(z.1, z.2, para$mu, para$sigma, para$rho, para$p) loglik.trace[j] <- l.new.outer if(j == 1) to.run <- T else{ # stop when iteration>100 if(j > 100) to.run <- F else to.run <- l.new.outer - l.old.outer > eps } # if(j %% 10==0) cat("loglik.trace[", j, "]=", loglik.trace[j], "\n") cat("mu=", para$mu, "sigma=", para$sigma, "p=", para$p, "rho=", para$rho, "\n") j <- j+1 } bic <- -2*l.new + 4*log(length(z.1)) return(list(para=list(p=para$p, rho=para$rho, mu=para$mu, sigma=para$sigma), loglik=l.new, bic=bic, e.z=e.z, loglik.trace=loglik.trace)) } # compute log-likelihood for mixture of two bivariate normals loglik.2binormal <- function(z.1, z.2, mu, sigma, rho, p){ l.m <- sum(d.binormal(z.1, z.2, 0, 1, 0)+log(p*exp(d.binormal(z.1, z.2, mu, sigma, rho)-d.binormal(z.1, z.2, 0, 1, 0))+(1-p))) # l.m <- sum((p*d.binormal(z.1, z.2, mu, sigma, rho) + (1-p)*d.binormal(z.1, z.2, 0, 1, 0))) return(l.m) } # check this when rho=1 # density of binomial distribution with equal mean and sigma on both dimensions d.binormal <- function(z.1, z.2, mu, sigma, rho){ loglik <- (-log(2)-log(pi)-2*log(sigma) - log(1-rho^2)/2 - (0.5/(1-rho^2)/sigma^2)*((z.1-mu)^2 -2*rho*(z.1-mu)*(z.2-mu) + (z.2-mu)^2)) return(loglik) } # E-step for computing the latent strucutre # e.z is the prob to be in the consistent group # e.step for estimating posterior prob # z.1 and z.2 can be vectors or scalars e.step.2normal <- function(z.1, z.2, mu, sigma, rho, p){ e.z <- p/((1-p)*exp(d.binormal(z.1, z.2, 0, 1, 0)-d.binormal(z.1, z.2, mu, sigma, rho))+ p) invisible(e.z) } # M-step for computing the latent structure # m.step for estimating proportion, mean, sd and correlation coefficient m.step.2normal <- function(z.1, z.2, e.z){ p <- mean(e.z) mu <- sum((z.1+z.2)*e.z)/2/sum(e.z) sigma <- sqrt(sum(e.z*((z.1-mu)^2+(z.2-mu)^2))/2/sum(e.z)) rho <- 2*sum(e.z*(z.1-mu)*(z.2-mu))/(sum(e.z*((z.1-mu)^2+(z.2-mu)^2))) return(list(p=p, mu=mu, sigma=sigma, rho=rho)) } # assume top p percent of observations are true # x and y are ranks, estimate init <- function(x, y, x.label){ x.o <- order(x) x.ordered <- x[x.o] y.ordered <- y[x.o] x.label.ordered <- x.label[x.o] n <- length(x) p <- sum(x.label)/n rho <- cor(x.ordered[1:ceiling(p*n)], y.ordered[1:ceiling(p*n)]) temp <- find.mu.sigma(x.ordered, x.label.ordered) mu <- temp$mu sigma <- temp$sigma invisible(list(mu=mu, sigma=sigma, rho=rho, p=p)) } # find mu and sigma if the distributions of marginal ranks are known # take the medians of the two dist and map back to the original init.dist <- function(f0, f1){ # take the median in f0 index.median.0 <- which(f0$cdf>0.5)[1] q.0.small <- f0$cdf[index.median.0] # because f0 and f1 have the same bins q.1.small <- f1$cdf[index.median.0] # take the median in f1 index.median.1 <- which(f1$cdf>0.5)[1] q.0.big <- f0$cdf[index.median.1] # because f0 and f1 have the same bins q.1.big <- f1$cdf[index.median.1] # find pseudo value for x.middle[1] on normal(0,1) pseudo.small.0 <- qnorm(q.0.small, mean=0, sd=1) pseudo.small.1 <- qnorm(q.1.small, mean=0, sd=1) # find pseudo value for x.middle[2] on normal(0,1) pseudo.big.0 <- qnorm(q.0.big, mean=0, sd=1) pseudo.big.1 <- qnorm(q.1.big, mean=0, sd=1) mu <- (pseudo.small.0*pseudo.big.1 - pseudo.small.1*pseudo.big.0)/(pseudo.big.1-pseudo.small.1) sigma <- (pseudo.small.0-mu)/pseudo.small.1 return(list(mu=mu, sigma=sigma)) } # generate labels # find the part of data with overlap # find the percentile on noise and signal # Suppose there are signal and noise components, with mean=0 and sd=1 for noise # x and x.label are the rank of the observations and their labels, # find the mean and sd of the other component # x.label takes values of 0 and 1 find.mu.sigma <- function(x, x.label){ x.0 <- x[x.label==0] x.1 <- x[x.label==1] n.x0 <- length(x.0) n.x1 <- length(x.1) x.end <- c(min(x.0), min(x.1), max(x.0), max(x.1)) o <- order(x.end) x.middle <- x.end[o][c(2,3)] # the smaller end of the overlap q.1.small <- mean(x.1 <= x.middle[1])*n.x1/(n.x1+1) q.0.small <- mean(x.0 <= x.middle[1])*n.x0/(n.x0+1) # the bigger end of the overlap q.1.big <- mean(x.1 <= x.middle[2])*n.x1/(n.x1+1) q.0.big <- mean(x.0 <= x.middle[2])*n.x0/(n.x0+1) # find pseudo value for x.middle[1] on normal(0,1) pseudo.small.0 <- qnorm(q.0.small, mean=0, sd=1) pseudo.small.1 <- qnorm(q.1.small, mean=0, sd=1) # find pseudo value for x.middle[2] on normal(0,1) pseudo.big.0 <- qnorm(q.0.big, mean=0, sd=1) pseudo.big.1 <- qnorm(q.1.big, mean=0, sd=1) mu <- (pseudo.small.0*pseudo.big.1 - pseudo.small.1*pseudo.big.0)/(pseudo.big.1-pseudo.small.1) sigma <- (pseudo.small.0-mu)/pseudo.small.1 return(list(mu=mu, sigma=sigma)) }