view intersectSig.py @ 12:2f4ea569f048

Uploaded
author xuebing
date Sat, 10 Mar 2012 08:10:44 -0500
parents b7f1d9f8f3bc
children
line wrap: on
line source

'''
find overlap and test signifiance
'''

import os,sys

def lineCount(filename):
    if os.stat(filename).st_size == 0:
        return 0
    with open(filename) as f:
        for i, l in enumerate(f):
            pass
            print i
    return i+1

def intersect(fileA,fileB,outfile,fraction,reciprocal):
    # return fileA intervals that overlap with interval in fileB
    cmd = 'intersectBed -a '+fileA+' -b '+fileB + ' -u -wa -f '+fraction +' '+ reciprocal + '>'+outfile
    #print cmd
    os.system(cmd)
    
def shuffle(fileA,fileB,genomefile,fraction,reciprocal,N):
    # shuffle fileA N times, return the distribution of overlaps
    nOverlap = []
    for i in range(N):
        # shuffle fileA using shuffleBed
        #cmd = 'shuffleBed -i '+fileA+' -g '+genomefile +'>fileA.shuffled'
        # using random_interval.py
        cmd = 'python /Users/xuebing/galaxy-dist/tools/mytools/random_interval.py '+fileA+' fileA.shuffled across '+genomefile
        os.system(cmd)
        intersect('fileA.shuffled',fileB,'tmp',fraction,reciprocal)
        nOverlap.append(lineCount('tmp'))
    os.system('rm tmp')
    os.system('rm fileA.shuffled')
    return nOverlap

def main():
    fileA = sys.argv[1]
    fileB = sys.argv[2]
    outfile = sys.argv[3]
    outplot = sys.argv[4]
    outshuffle = sys.argv[5]
    N = int(sys.argv[6]) # times to shuffle
    genomefile = sys.argv[7]
    fraction = sys.argv[8]
    if len(sys.argv) == 10:
        reciprocal = sys.argv[9] # can only be '-r'
    else:
        reciprocal = ''

    #print sys.argv

    # number of lines in input
    nA = lineCount(fileA)
    nB = lineCount(fileB)    

    # intersect on real data
    intersect(fileA,fileB,outfile,fraction,reciprocal)
    # number of overlaps
    nOverlapReal = lineCount(outfile)

    #print 'number of intervals in inputA that overlap with intervals in inputB:',nOverlapReal
    
    # shuffle fileA to estimate background
    nOverlapNull = shuffle(fileA,fileB,genomefile,fraction,reciprocal,N)
    out = open(outshuffle,'w')
    out.write("\t".join(map(str,nOverlapNull)))
    out.close()

    # plot histogram
    rscript = open('tmp.r','w')
    rscript.write("options(warn=-1)\n")
    rscript.write("x0 <- "+str(nOverlapReal)+"\n")
    rscript.write("x <- c("+','.join(map(str,nOverlapNull))+")\n")
    rscript.write("library(MASS)\n")
    rscript.write("pv <- min((1+sum(x>=x0))/length(x),(1+sum(x<=x0))/length(x))\n")
    rscript.write("title <- paste('actual:chance = ',x0,':',format(mean(x),digits=1,nsmall=1),' = ',format(x0/mean(x),digits=1,nsmall=2),', p-value < ',pv,sep='')\n")
    rscript.write("pdf('"+outplot+"')\n")
    rscript.write("library(grid)\n")
    rscript.write("library(VennDiagram)\n")
    rscript.write("venn <- venn.diagram(x=list(A=1:"+str(nA)+",B="+str(nA-nOverlapReal+1)+":"+str(nA+nB-nOverlapReal)+"),filename=NULL,fill=c('red','blue'),col='transparent',alpha=0.5,label.col='black',cex=3,lwd=0,fontfamily='serif',fontface='bold',cat.col = c('red', 'blue'),cat.cex=3,cat.fontfamily='serif',cat.fontface='bold')\n")
    rscript.write("grid.draw(venn)\n")
    rscript.write("h <- hist(x,breaks=50,xlab='number of overlaps',ylab='frequency',main=title)\n")
    rscript.write("plot(h$mids,h$counts,type='h',xlim=c(min(h$mids,x0),max(x0,h$mids)),ylim=c(0,max(h$counts)),xlab='number of overlaps',ylab='frequency',main=title)\n")
    rscript.write("points(x0,0,col='red')\n")
    rscript.write("dev.off()\n")
    rscript.close()
    os.system("R --vanilla < tmp.r")    
    os.system('rm tmp.r')
main()