Introducing KmerCompressor, a tool for set operations on kmers

[luc]

Very neat! Time to start playing.

[Brian Bushnell]

It's currently capped at 31, though I could make an unlimited-kmer-length version in a few hours. That would probably be worth doing, if I get some free time.

[GenoMax]

What is the upper limit on the k-mer size one can specify?

[Brian Bushnell]

I'd like to introduce a new BBTool, KmerCompressor. This will take a dataset and reduce it to its set of constituent kmers, and print an optimally-condensed representation of them in fasta format, in which each kmer occurs exactly once. This is similar to an assembler, but it has additional capabilities regarding kmer count cutoffs that allow it to be used to perform arbitrary set operations on kmers, which allows advanced filtering of raw reads to capture specific features such as ribosomes, mitochondria, and chloroplasts, or filter by taxonomy.

The basic usage is like this:
kcompress.sh in=reads.fq out=set.fa

To get just the 31-mers that appear between 100 and 150 times in a dataset:
kcompress.sh in=reads.fq out=set.fa min=100 max=150 k=31

To use it for a set union (all the kmers in either of two files):
kcompress.sh in=ecoli.fa,salmonella.fa out=union.fa

With those basic operations, it is now possible to do various set operations. For example:
kcompress.sh in=fungal_genome.fa out=set_g.fa
kcompress.sh in=fungal_mitochondria.fa out=set_m.fa

Each of those sets has each kmer represented exactly once. Therefore, you can perform an intersection like this:
kcompress.sh in=set_g.fa,set_m.fa out=intersection.fa min=2

Or a subtraction like this:
kcompress.sh in=set_m.fa,intersection.fa out=m_minus_g.fa max=1

Then m_minus_g.fa contains all the kmers that are specific only to mitochondria in that organism, and could be used for filtering reads in an iterative assembly process.
I've been recently using it to create a set of ribosomal kmers for rapid metatranscriptome rRNA filtering using BBDuk, by reducing a very large ribosomal (16S/18S) database to just the set of kmers that occur often (and are thus both correct and conserved). This is useful for avoiding false positives, and reducing load time and memory usage compared to working with the entire database. For example:
dedupe.sh in=multiple_ribo_databases.fa.gz out=nodupes.fa.gz
kcompress.sh in=nodupes.fa.gz out=compressed.fa.gz k=31 min=5

...will result in a much smaller file, with similar (tunable) sensitivity and better specificity compared to the original. Subsequently, I run:
bbduk.sh in=metatranscriptome.fq.gz outu=nonribo.fq.gz outm=ribo.fq.gz ref=compressed.fa.gz k=31

...to separate the reads.

P.S. A link to a file I created with KmerCompressor: ribokmers.fa.gz
This 9MB file contains commonly-occurring ribosomal kmers from Silva. Used in conjunction with BBDuk, like this:

bbduk.sh in=reads.fq outm=ribo.fq outu=nonribo.fq k=31 ref=ribokmers.fa.gz

...it has a roughly 99.94% sensitivity against synthetic 1x150bp from the full Silva database (180MB compressed), a 99.98% sensitivity with hdist=1, and 99.994% sensitivity at k=25 hdist=1.