抱歉,您的浏览器无法访问本站
本页面需要浏览器支持(启用)JavaScript
了解详情 >

拆分原理

  • 软件的逻辑是首先获取barcode列表。然后采用多线程分别在fastq文件中并行提取对应barcode的reads。

  • WGS的下机数据经常出现在fastq2里。所以程序会从fastq中自动查找是否存在对应barcode。

  • 程序可以自动检测barcode始于开始还是末尾,计算hanming距离,运行1bp的mismatch。

  • 程序可以自动检测barcode始于开始还是末尾,计算hanming距离,运行1bp的mismatch。

方法

julia -t number threads split_barcode.jl -b barcode list -r read2 file -l read1 file -o outpath

代码举例

using BioSequences
using FASTX
using CodecZlib
using StringDistances
using Base.Threads
using Getopt

function read_barcode(barcode_fs::String)
labels = String[]
barcodes = String[]
if isfile(barcode_fs)
for line in eachline(barcode_fs)
label, sequence = split(line)
push!(labels, label)
push!(barcodes, sequence)
end
else
println("no barcode file detected!")
end
labels, barcodes
end

function detective_barcode(seq::LongDNASeq, barcode::LongDNASeq)
barcode_l = length(barcode)
starter = seq[1:barcode_l]
ender = seq[end-barcode_l:end]
dist = Hamming()(starter, barcode)
if dist == 0 || dist == 1 ## if barcode in title
return 1, 1
else
dist = Hamming()(ender, barcode)# calular hamming distance allow 1bp mismatch
if dist == 0 || dist == 1
return 1, 2
else
return -1, 0
end
end
end

function process(fq_file_1::String, fq_file_2::String, barcode::LongDNASeq, label::String, outpath::String)
reader_2 = FASTQ.Reader(GzipDecompressorStream(open(fq_file_2)))
reader_1 = FASTQ.Reader(GzipDecompressorStream(open(fq_file_1)))
wt_1 = FASTQ.Writer(open("$(outpath)/barcode_$(label)_1.fq", lock=true, "a")) ## lock for multiple threads safety
wt_2 = FASTQ.Writer(open("$(outpath)/barcode_$(label)_2.fq", lock=true, "a"))
for (record_1, record_2) in zip(reader_1, reader_2)
fm, weizhi = detective_barcode(FASTQ.sequence(record_2), barcode)
if fm == 1
write(wt_1, record_1)
if weizhi == 1
write(wt_2, FASTQ.Record(FASTQ.identifier(record_2), FASTQ.sequence(record_2)[11:end], FASTQ.quality(record_2)[11:end]))
elseif weizhi == 2
write(wt_2, FASTQ.Record(FASTQ.identifier(record_2), FASTQ.sequence(record_2)[1:100], FASTQ.quality(record_2)[1:100]))
end
end
end
close(reader_1)
close(reader_2)
close(wt_1)
close(wt_2)
end


function split_barcode(barcodes_fs::String, fq_file_2::String, fq_file_1::String, outpath::String)
labels, barcodes = read_barcode(barcodes_fs)
barcodes = LongDNASeq.(barcodes)
duiying = Dict(zip(barcodes, labels))
@threads for barcode in barcodes
label = duiying[barcode]
process(fq_file_1, fq_file_2, barcode, label, outpath)
end
end

function Argparse()
lst = Dict{String,String}()
for (opt, arg) in getopt(ARGS, "hb:r:l:o:", ["help","barcodes=", "read2=","read1=","output="])
opt = replace(opt, "-" => "")
arg = replace(arg, "=" => "")
if opt == "help" || opt == "h"
println("this program is for spliting fastq file into different part according barcodes!")
println("-h\t--help\tshow this message")
println("-b\t--barcodes\tinput your barcodes file, should be negetive sequennce!")
println("-r2\t--read2\tinput your fastq file read_2")
println("-r1\t--read1\tinput your fastq file read_1")
println("-o\t--output\tinput output file path")
elseif opt == "barcodes" || opt == "b"
lst["barcodes"] = arg
elseif opt == "read2" || opt == "r"
lst["read2"] = arg
elseif opt == "read1" || opt == "l"
lst["read1"] = arg
elseif opt == "output" || opt == "o"
lst["output"] = arg
else
println("please check your parameter!")
end
end
lst
end


args = Argparse()
if length(args) > 1
split_barcode(args["barcodes"], args["read2"], args["read1"], args["output"])
end

评论