Mercurial > repos > xuebing > sharplabtool
view tools/samtools/pileup_parser.pl @ 2:c2a356708570
Uploaded
author | xuebing |
---|---|
date | Fri, 09 Mar 2012 19:45:42 -0500 |
parents | 9071e359b9a3 |
children |
line wrap: on
line source
#! /usr/bin/perl -w use strict; use POSIX; die "Usage: pileup_parser.pl <in_file> <ref_base_column> <read_bases_column> <base_quality_column> <coverage column> <qv cutoff> <coverage cutoff> <SNPs only?> <output bed?> <coord_column> <out_file> <total_diff> <print_qual_bases>\n" unless @ARGV == 13; my $in_file = $ARGV[0]; my $ref_base_column = $ARGV[1]-1; # 1 based my $read_bases_column = $ARGV[2]-1; # 1 based my $base_quality_column = $ARGV[3]-1; # 1 based my $cvrg_column = $ARGV[4]-1; # 1 based my $quality_cutoff = $ARGV[5]; # phred scale integer my $cvrg_cutoff = $ARGV[6]; # unsigned integer my $SNPs_only = $ARGV[7]; # set to "Yes" to print only positions with SNPs; set to "No" to pring everything my $bed = $ARGV[8]; #set to "Yes" to convert coordinates to bed format (0-based start, 1-based end); set to "No" to leave as is my $coord_column = $ARGV[9]-1; #1 based my $out_file = $ARGV[10]; my $total_diff = $ARGV[11]; # set to "Yes" to print total number of deviant based my $print_qual_bases = $ARGV[12]; #set to "Yes" to print quality and read base columns my $invalid_line_counter = 0; my $first_skipped_line = ""; my %SNPs = ('A',0,'T',0,'C',0,'G',0); my $above_qv_bases = 0; my $SNPs_exist = 0; my $out_string = ""; my $diff_count = 0; open (IN, "<$in_file") or die "Cannot open $in_file $!\n"; open (OUT, ">$out_file") or die "Cannot open $out_file $!\n"; while (<IN>) { chop; next if m/^\#/; my @fields = split /\t/; next if $fields[ $ref_base_column ] eq "*"; # skip indel lines my $read_bases = $fields[ $read_bases_column ]; die "Coverage column" . ($cvrg_column+1) . " contains non-numeric values. Check your input parameters as well as format of input dataset." if ( not isdigit $fields[ $cvrg_column ] ); next if $fields[ $cvrg_column ] < $cvrg_cutoff; my $base_quality = $fields[ $base_quality_column ]; if ($read_bases =~ m/[\$\^\+-]/) { $read_bases =~ s/\^.//g; #removing the start of the read segement mark $read_bases =~ s/\$//g; #removing end of the read segment mark while ($read_bases =~ m/[\+-]{1}(\d+)/g) { my $indel_len = $1; $read_bases =~ s/[\+-]{1}$indel_len.{$indel_len}//; # remove indel info from read base field } } if ( length($read_bases) != length($base_quality) ) { $first_skipped_line = $. if $first_skipped_line eq ""; ++$invalid_line_counter; next; } # after removing read block and indel data the length of read_base # field should identical to the length of base_quality field my @bases = split //, $read_bases; my @qv = split //, $base_quality; for my $base ( 0 .. @bases - 1 ) { if ( ord( $qv[ $base ] ) - 33 >= $quality_cutoff and $bases[ $base ] ne '*') { ++$above_qv_bases; if ( $bases[ $base ] =~ m/[ATGC]/i ) { $SNPs_exist = 1; $SNPs{ uc( $bases[ $base ] ) } += 1; $diff_count += 1; } elsif ( $bases[ $base ] =~ m/[\.,]/ ) { $SNPs{ uc( $fields[ $ref_base_column ] ) } += 1; } } } if ($bed eq "Yes") { my $start = $fields[ $coord_column ] - 1; my $end = $fields[ $coord_column ]; $fields[ $coord_column ] = "$start\t$end"; } if ($print_qual_bases ne "Yes") { $fields[ $base_quality_column ] = ""; $fields[ $read_bases_column ] = ""; } $out_string = join("\t", @fields); # \t$read_bases\t$base_quality"; foreach my $SNP (sort keys %SNPs) { $out_string .= "\t$SNPs{$SNP}"; } if ($total_diff eq "Yes") { $out_string .= "\t$above_qv_bases\t$diff_count\n"; } else { $out_string .= "\t$above_qv_bases\n"; } $out_string =~ s/\t+/\t/g; if ( $SNPs_only eq "Yes" ) { print OUT $out_string if $SNPs_exist == 1; } else { print OUT $out_string; } %SNPs = (); %SNPs = ('A',0,'T',0,'C',0,'G',0); $above_qv_bases = 0; $SNPs_exist = 0; $diff_count = 0; } print "Skipped $invalid_line_counter invalid line(s) beginning with line $first_skipped_line\n" if $invalid_line_counter > 0; close IN; close OUT;