Genome STRiP makes use of mask files that identify portions of the reference sequence that are not reliably alignable.
Genome mask files are fasta files with the same number of sequences and of the same length as the reference sequence. In a genome mask file, a base position is marked with a 0 if it is reliably alignable and 1 if it is not. Each genome mask file is specific to the reference sequence and to the parameters used to determine alignability.
The current generation of mask files are based on fixed read lengths. A base is assigned a 0 if an N base sequence centered on this read is unique within the reference genome. You should use a genome mask with a value of N that corresponds to the read lengths of your input data set. For example, if you have data that is a uniform set of Illumina paired-end data with 101bp reads, then you should use (or generate) a genome mask with a read length of 101. If your data is a mixture of read lengths, one viable strategy is to use a "lowest common denominator" approach and use a mask length corresponding to the shortest reads in your input data set. Using the smallest read length will cause a small additional fraction of the genome to be marked inaccessible, but will give the best specificity. Alternatively, you can use a larger N, which should modestly improve sensitivity at the cost of a modest increase in false discovery rate and a modest decrease in genotyping accuracy.
Some precomputed mask files for a variety of reference sequences and read lengths are available at ftp://ftp.broadinstitute.org/pub/svtoolkit/svmasks.
The ComputeGenomeMask command line utility is available to generate genome mask files, but queue scripts to automate the process have not been written. A reasonable strategy is to compute the genome mask in parallel chromsome-by-chromosome and then merge the resulting fasta files into a final genome-wide mask file.
The implementation of mask files will be replaced in a future release.
Mask files are being converted from textual fasta files to binary files and are being enhanced to better support input data sets with multiple read lengths (so the use of a "lowest common denominator" strategy will no longer be necessary).