dashing sketches and computes distances between fasta and fastq data.
Our paper is available here as a preprint and here at Genome Biology.
The easiest way to use dashing is to download a binary release. These are located on the release page of the dashing-binaries repo. These archives contain three versions per executable, called dashing_s128, dashing_s256, and dashing_s512. These work, respectively, on systems supporting SSE2, AVX2, and AVX512BW instructions. If a binary with a higher number after the s fails to work on your system, try one with a lower number. Note also that the binaries are gzipped, so you may need to gunzip * in the appropriate binary directory first.
Clone this repository recursively, and use make.
git clone --recursive https://github.com/dnbaker/dashing
cd dashing && make dashingIf you clone without submodules, this can be corrected by git submodule update --init --recursive.
Dashing is written in C++14, which means that it requires a relatively new compiler.
Dashing is tested under gcc{5.4-9}, but fails for gcc4, which is installed by default on many machines.
For OSX, we recommend using Homebrew to install gcc-8.
On Linux, we recommend package managers. (For instance, our Travis-CI Ubuntu example upgrades to a sufficiently new GCC using sudo update-alternatives.)
To see all usage options, use ./dashing <subcommand>, for subcommand in [sketch, dist, hll, union, printmat].
Of most interest is probably the dist command, which can take either genomes or pre-built sketches as arguments.
For the simplest case of unspaced, unminimized kmers for a set of genomes with k = 31 and 13 threads:
dashing dist -k31 -p13 -Odistance_matrix.txt -osize_estimates.txt genome1.fna.gz genome2.fna genome3.fasta <...>
The genomes can be omitted as positional arguments if -F genome_paths.txt is provided, where genome_paths.txt is a file containing a path to a genome per line.
This can avoid system limits on the number of arguments in a shell command.
These can be cached with -c, which saves the sketches for later use. These sketch filenames are based on spacing, kmer size, and sketch size, so there is no risk of overwriting each other.
dashing dist performs all pairwise jaccard index estimates by default. By providing the -Q flag, dashing performs a core
comparison operation between all queries and all references, where references are provided by -F.
This is necessary to provide containment.
For example:
dashing dist --containment-index -k21 -Odistmat.txt -ofsizes.txt -Q query_paths.txt -F ref_paths.txt
To generate a full, asymmetric distance matrix, provide the same path to -F and -Q.
The sketch command largely mirrors dist, except that only sketches are computed.
dashing sketch -k31 -p13 -F genome_paths.txt
The hll command simply estimates the number of unique elements in a set of files. This can be useful for estimating downstream database sizes based on spacing schemes.
The union command takes a set of pre-sketched HLLs and performs unions between them. Currently, the sketches must be of the same size. We may modify this in future releases to allow a merger of different sizes by flattening larger sketches to the smallest sketch size discovered. This would involve a loss of precision from the larger models. This currently doesn't support data structures besides HLLs, but we plan to make this change at a later date.
Transparently consumes uncompressed, zlib- or zstd-compressed files.
Caching of sketches to disk (in compressed form)
Calculation of a variety of (dis)similarity measures:
- Jaccard Similarity
- Mash distance
- Containment index
- Containment distance (log transformed containment index)
- Symmetric Containment Index (
\frac{|A \bigcap B|}{\min{|A|,|B|}}) (The maximum of each containment index) - Symmetric Continament Distance (log transformed SCI)
- Intersection size
Additionally, supports all the above under the weighted/multiset Jaccard index via labeled w-shingling. (See Broder, 1997 "On the Resemblance and Containment of Documents" for more details.)
Filtering of of rare k-mer events via count-min sketch point query estimates. This is primarily desirable for raw sequencing datasets rather than genome assemblies. This is enabled with -y/--countmin, and the number of hashes (--nhashes), sketch size (--cm-sketch-size) and min count --min-count can all be controlled by command-line parameters.
- Exact k-mer encoding (
k <= 32) - Rolling hashing encoding for any k
- Spaced seed encoding (Hamming weight <= 32)
- Windowed/minimized k-mers
See the bonsai for more details on encoding.
Dashing defaults to upper triangular TSV matrix emission, but it also suppurts upper triangular PHYLIP format, packed binary encoding, and top k-nearest-neighbor emission formats.
This is supported for all symmetric measures (Mash distance, Jaccard, Intersection size, and their multiset equivalents), whereas asymmetric measures and nearest neighbor forms (all variations of containment) have two emission options: tabular and binary.
Dashing supports comparisons with a variety of data structures, which have speed and accuracy tradeoffs for given situations. By default, HyperLogLog sketches are used, while b-bit minhashing, bottom-k minhashing, bloom filters, and hash sets are supported. Using hash sets provides a ground truth at the expense of greatly increased runtime costs.
b-bit minhashing: --use-bb-minhash
bottom-k minhashing: --use-range-minhash
hash sets: --use-full-khash-sets
bloom filters: --use-bloom-filter
Wide HLL: --use-wide-hll
References: SuperMinHash, modified. (Use 32-bit register instead of float between 0 and 1 to make use of more information.) Bloom Filter Jaccard Index
@Article{pmid31801633,
Author="Baker, D. N. and Langmead, B. ",
Title="{{D}ashing: fast and accurate genomic distances with {H}yper{L}og{L}og}",
Journal="Genome Biol.",
Year="2019",
Volume="20",
Number="1",
Pages="265",
Month="12"
}
