sourmash Python API examples

A first example: two k-mers

Define two sequences:

>>> seq1 = "ATGGCA"
>>> seq2 = "AGAGCA"

Create two MinHashes using 3-mers, and add the sequences:

>>> import sourmash
>>> E1 = sourmash.MinHash(n=20, ksize=3)
>>> E1.add_sequence(seq1)
>>> E2 = sourmash.MinHash(n=20, ksize=3)
>>> E2.add_sequence(seq2)

One of the 3-mers (out of 7) overlaps, so Jaccard index is 1/7:

>>> round(E1.jaccard(E2), 2)
0.14

and of course the MinHashes match themselves:

>>> E1.jaccard(E1)
1.0

We can add sequences and query at any time –

>>> E1.add_sequence(seq2)
>>> x = E1.jaccard(E2)
>>> round(x, 3)
0.571

Consuming files

Suppose we want to create MinHash sketches from genomes –

>>> import glob, pprint
>>> genomes = glob.glob('data/GCF*.fna.gz')
>>> genomes = list(sorted(genomes))
>>> pprint.pprint(genomes)
['data/GCF_000005845.2_ASM584v2_genomic.fna.gz',
 'data/GCF_000006945.1_ASM694v1_genomic.fna.gz',
 'data/GCF_000783305.1_ASM78330v1_genomic.fna.gz']

We have to read them in (here using screed), but then they can be fed into add_sequence directly; here we set force=True in add_sequence to skip over k-mers containing characters other than ACTG, rather than raising an exception.

(Note, just for speed reasons, we’ll truncate the sequences to 50kb in length.)

>>> import screed
>>> minhashes = []
>>> for g in genomes:
...     E = sourmash.MinHash(n=500, ksize=31)
...     for record in screed.open(g):
...         E.add_sequence(record.sequence[:50000], True)
...     minhashes.append(E)

And now the minhashes can be compared against each other:

>>> import sys
>>> for i, e in enumerate(minhashes):
...    _ = sys.stdout.write(genomes[i][:20] + ' ')
...    for j, e2 in enumerate(minhashes):
...       x = e.jaccard(minhashes[j])
...       _ = sys.stdout.write(str(round(x, 3)) + ' ')
...    _= sys.stdout.write('\n')
data/GCF_000005845.2 1.0 0.0 0.0 
data/GCF_000006945.1 0.0 1.0 0.0 
data/GCF_000783305.1 0.0 0.0 1.0 

Note that the comparisons are quite quick; most of the time is spent in making the minhashes, which can be saved and loaded easily.

Saving and loading signature files

Signature files encapsulate MinHashes in JSON, and provide a way to add some metadata to MinHashes.

>>> from sourmash import SourmashSignature, save_signatures
>>> sig1 = SourmashSignature(minhashes[0], name=genomes[0][:20])
>>> with open('/tmp/genome1.sig', 'wt') as fp:
...   save_signatures([sig1], fp)

Here, /tmp/genome1.sig is a JSON file that can now be loaded and compared – first, load:

>>> from sourmash import load_one_signature
>>> loaded_sig = load_one_signature('/tmp/genome1.sig')

then compare:

>>> loaded_sig.jaccard(sig1)
1.0
>>> sig1.jaccard(loaded_sig)
1.0

Manipulating signatures and their hashes.

It is relatively straightforward to work directly with hashes.

First, load two signatures:

>>> sigfile1 = 'tests/test-data/genome-s10.fa.gz.sig'
>>> sig1 = load_one_signature(sigfile1, ksize=21, select_moltype='DNA')
>>> sigfile2 = 'tests/test-data/genome-s11.fa.gz.sig'
>>> sig2 = load_one_signature(sigfile2, ksize=21, select_moltype='DNA')

Then, get the hashes, and (e.g.) compute the union:

>>> hashes1 = set(sig1.minhash.get_mins())
>>> hashes2 = set(sig2.minhash.get_mins())
>>> hash_union = hashes1.union(hashes2)
>>> print('{} hashes in union of {} and {}'.format(len(hash_union), len(hashes1), len(hashes2)))
1000 hashes in union of 500 and 500

sourmash MinHash objects and manipulations

sourmash supports two basic kinds of signatures, MinHash and modulo hash signatures. MinHash signatures are equivalent to mash signatures; they are limited in size, and very effective for comparing genomes and other data sets that are of similar size. The key parameter for MinHash signatures is num, which specifies the maximum number of hashes to be collected for a given input data set.

>>> signum = sourmash.MinHash(n=500, ksize=31)

Because of this parameter, below we’ll call them ‘num’ signatures.

Modulo hash (or ‘scaled’) signatures are specific to sourmash and they enable an expanded range of metagenome analyses, with the downside that they can become arbitrarily large. The key parameter for modulo hash signatures is scaled, which specifies the average sampling rate for hashes for a given input data set. A scaled factor of 1000 means that, on average, 1 in 1000 k-mers will be turned into a hash for later comparisons; this is a sort of compression factor, in that a 5 Mbp genome will yield approximately 5000 hash values with a scaled factor of 1000 (5000 x 1000 = 5,000,000).

>>> sigscaled = sourmash.MinHash(n=0, ksize=31, scaled=1000)

Note also that with a scaled factor of 1, the signature will contain all of the k-mers.


You can differentiate between num signatures and scaled signatures by looking at the num and scaled attributes on a MinHash object:

>>> signum.num
500
>>> signum.scaled
0
>>> sigscaled.num
0
>>> sigscaled.scaled
1000

The MinHash class is otherwise identical between the two types of signatures.

Note that you cannot compute Jaccard similarity or containment for MinHash objects with different num or scaled values (or different ksizes):

>>> signum2 = sourmash.MinHash(n=1000, ksize=31)
>>> signum.similarity(signum2)
Traceback (most recent call last):
  ...
TypeError: must have same num: 500 != 1000

You can make signatures compatible by downsampling; see the next sections.

A brief introduction to MinHash object methods and attributes

MinHash objects have the following methods and attributes:

  • ksize, num, and scaled - the basic parameters used to create a MinHash object.
  • get_mins() - retrieve all of the hashes contained in this object.
  • add_sequence(seq) - hash sequence and add hash values.
  • add(hash) and add_many(hashvals) - add hash values directly.
  • similarity(other) - calculate Jaccard similarity with the other MinHash object.
  • contained_by(other) - calculate the Jaccard containment of self by other.
  • copy_and_clear() - make an empty copy of a MinHash object with the same parameters.
  • __len__() - return the number of actual hash values. Note you can also do len(mh), where mh is a MinHash object.

Downsampling signatures

Num and scaled signatures can always be downsampled without referring back to the original data.

Let’s start by loading 50kb of genomic sequence in to memory:

>>> genomes = glob.glob('data/GCF*.fna.gz')
>>> genomes = list(sorted(genomes))
>>> genome = genomes[0]
>>> record = next(iter(screed.open(genome)))
>>> sequence = record.sequence[:50000]

Now, suppose we make a num signature limited to 1000 hashes:

>>> larger = sourmash.MinHash(n=1000, ksize=31)
>>> larger.add_sequence(sequence)
>>> len(larger)
1000

We can downsample this to 500 by extracting the hashes and using add_many to add them to a new MinHash like so:

>>> hashvals = larger.get_mins()
>>> smaller = sourmash.MinHash(n=500, ksize=31)
>>> smaller.add_many(hashvals)
>>> len(smaller)
500

Also note that there’s a convenience function that does the same thing, faster!

>>> smaller2 = larger.downsample_n(500)
>>> smaller2 == smaller
True

The same can be done with scaled MinHashes:

>>> large_scaled = sourmash.MinHash(n=0, ksize=31, scaled=100)
>>> large_scaled.add_sequence(sequence)
>>> len(large_scaled)
459
>>> small_scaled = sourmash.MinHash(n=0, ksize=31, scaled=500)
>>> small_scaled.add_many(large_scaled.get_mins())
>>> len(small_scaled)
69

And, again, there’s a convenience function that you can use:

>>> small_scaled2 = large_scaled.downsample_scaled(500)
>>> small_scaled == small_scaled2
True

Converting between num and scaled signatures

(Beware, these are confusing techniques for working with hashes that are easy to get wrong! We suggest posting questions in the issue tracker as you go, if you are interested in exploring this area!)

The hashing function used is identical between num and scaled signatures, so the hash values themselves are compatible - it’s the comparison between collections of them that doesn’t work.

But, in some circumstances, num signatures can be extracted from scaled signatures, and vice versa. We haven’t yet implemented a Python API for this in sourmash, but you can hack it together yourself quite easily, and a conversion utility is implemented through the command line in sourmash signature downsample.

To extract a num MinHash object from a scaled MinHash, first create or load your MinHash, and then extract the hash values:

>>> num_mh = sourmash.MinHash(n=1000, ksize=31)
>>> num_mh.add_sequence(sequence)
>>> hashvals = num_mh.get_mins()

Now, create the new scaled MinHash object and add the hashes to it:

>>> scaled_mh = sourmash.MinHash(n=0, ksize=31, scaled=10000)
>>> scaled_mh.add_many(hashvals)

and you are done!

The same works in reverse, of course:

>>> scaled_mh = sourmash.MinHash(n=0, ksize=31, scaled=50)
>>> scaled_mh.add_sequence(sequence)
>>> hashvals = scaled_mh.get_mins()
>>> num_mh = sourmash.MinHash(n=500, ksize=31)
>>> num_mh.add_many(hashvals)

So… when can you do this extraction reliably?

You can extract num MinHashes from scaled MinHashes whenever the maximum hash value in the num MinHash is greater than or equal to the max_hash attribute of the scaled MinHash.

You can extract scaled MinHashes to num MinHashes whenever there are more hash values in the scaled MinHash than num.

Yoda sayeth: When understand these two sentences you can, use this code may you.

(You can also take a look at the logic in sourmash signature downsample if you are interested.)

Working with fast search trees (Sequence Bloom Trees, or SBTs)

Suppose we have a number of signatures calculated with --scaled, like so:

sourmash compute --scaled 10000 data/GCF*.fna.gz

and now we want to create a Sequence Bloom Tree (SBT) so that we can search them efficiently. You can do this with sourmash index, but you can also access the Python API directly.

Creating a search tree

Let’s start by using ‘glob’ to grab some example signatures from the test data in the sourmash repository:

>>> import glob
>>> input_filenames = glob.glob('tests/test-data/doctest-data/GCF*.sig')

Now, create a tree:

>>> import sourmash
>>> tree = sourmash.create_sbt_index()

Load each signature, and add it to the tree:

>>> from sourmash.sbtmh import SigLeaf
>>> for filename in input_filenames:
...     sig = sourmash.load_one_signature(filename, ksize=31)
...     leaf = SigLeaf(sig.md5sum(), sig)
...     tree.add_node(leaf)

(note, you’ll need to make sure that all of the signatures are compatible with each other! The sourmash index command does all of the necessary checks.)

Now, save the tree:

>>> filename = tree.save('/tmp/test.sbt.json')

Loading and search SBTs

How do we load the SBT and search it with a DNA sequence, from within Python?

The SBT filename is /tmp/test.sbt.json, as above:

>>> SBT_filename = '/tmp/test.sbt.json'

and with it we can load the SBT:

>>> tree = sourmash.load_sbt_index(SBT_filename)

Now, load a DNA sequence:

>>> filename = 'data/GCF_000005845.2_ASM584v2_genomic.fna.gz'
>>> query_seq = next(iter(screed.open(filename))).sequence
>>> print('got {} DNA characters to query'.format(len(query_seq)))
got 4641652 DNA characters to query

and create a scaled signature:

>>> minhash = sourmash.MinHash(ksize=31, n=0, scaled=10000)
>>> minhash.add_sequence(query_seq)

>>> query_sig = sourmash.SourmashSignature(minhash, name='my favorite query')

Now do a search –

>>> threshold = 0.1
                                           
>>> for found_sig, similarity in sourmash.search_sbt_index(tree, query_sig, threshold):
...    print(query_sig.name())
...    print(found_sig.name())
...    print(similarity)
my favorite query
NC_000913.3 Escherichia coli str. K-12 substr. MG1655, complete genome
1.0

et voila!