fastar is a fasta region generator.
You use it to generate a list of genomic regions with even sequencing coverage. This can be used with variant callers to reduce the variance in calling runtime for each region.
The easiest way to install fastar is to download the latest binary from
here and then chmod +x the file.
To build from source you need a working go installation. Then just:
go install github.com/sb10/fastar
The command has help accessible by running with the -h option.
fastar -[r|b] <int> samtools.depth.output
It is recommended to use -b to set the desired total coverage (in base pairs) per region, eg. 2000000. Using -r instead results in it taking almost twice as long, and doesn't work with piped input, since this results in base pairs per region being calculated by summing the depths in the depth file, before going through it again to calculate the regions.
Note also that with -r, the number of regions you get isn't guaranteed. It may be much higher if you have coverage over many small sequences: the minimum number of regions will be the number of covered sequences in your reference.
fastar is a pure-go version of the 'coverage_to_regions.py' script supplied with freebayes.
fastar was written because the python version is slow and uses lots of memory: on this author's ~3billion line depth file, it consumed over 250GB of memory before crashing after 23hrs.
fastar, on the same file, takes ~12mins (~8mins if using -b) and consumes a constant small amount of memory. On a ~50million line truncated version of the file, coverage_to_regions.py took 308s (and used ~1GB), while fastar took 14s (and used 7MB), 9s with -b, making it about 22-34 times faster.
fastar differs from coverage_to_regions.py in that its regions start and end on the first and last base per reference sequence that had coverage, instead of starting at 0 (an invalid coordinate) and ending at the end of the sequence. Because of this you don't need to supply the reference index file. Regions output also do not overlap.
There is a tool in goleft called indexsplit that operates in seconds and doesn't require the depth file, but it isn't as accurate. However, typically accuracy isn't that important for this use case, and indexsplit may be good enough.
That said, what determines calling time is related to base pairs per region, so it is more useful to generate regions based on a fixed value for that, as opposed to a fixed number of regions, which is the only option available with goleft (and the original freebayes script).
fastar also only outputs regions that have coverage, so you don't get regions for whole references sequences where no alignments occurred. This makes the output more suitable for passing to callers, since you now won't end up wasting time running calls on zero data, and you don't have to deal with empty VCF files.