Classifying signatures: search, gather, and lca methods.

sourmash provides several different techniques for doing classification and breakdown of signatures.

Analyzing metagenomic samples with gather

Neither search option (similarity or containment) is effective when comparing or searching with metagenomes, which typically contain a mixture of many different genomes. While you might use containment to see if a query genome is present in one or more metagenomes, a common question to ask is the reverse: what genomes are in my metagenome? An alternative phrasing is this: what reference genomes should I map my metagenomic reads to?

The main approach we provide in sourmash is sourmash gather. This constructs the shortest possible list of reference genomes that cover all of the known k-mers in a metagenome. We call this a minimum metagenome cover.

From an algorithmic perspective, gather generates a minimum set cover for a query metagenome, using the reference database you give it. The minimum set cover is calculated using a greedy approximation algorithm. Essentially, gather takes a query metagenome and searches the database for the most highly contained genome; it then subtracts that match from the metagenome, and repeats. At the end it reports how much of the metagenome remains unknown. The basic sourmash tutorial has some sample output from using gather with GenBank. See Appendix A at the bottom of this page for more technical details.

The gather method is described in Lightweight compositional analysis of metagenomes with FracMinHash and minimum metagenome covers, Irber et al., 2022. Our benchmarking in that paper and also in Evaluation of taxonomic classification and profiling methods for long-read shotgun metagenomic sequencing datasets, Portik et al., 2022 suggests that it is a very sensitive and specific method for analyzing metagenomes.

Taxonomic profiling with sourmash

sourmash supports two basic kinds of taxonomic profiling, under the lca and tax modules. As of 2023, we strongly recommend the tax-based profiling approach.

But first, let’s back up! By default, there is no structured taxonomic information available in sourmash signatures or collections. What this means is that you will have to provide your own mapping from a match to some taxonomic hierarchy. Both the lca and tax modules support identifier-based taxonomic mappings, in which identifiers from the signature names can be linked to the standard seven NCBI/GTDB taxonomy ranks - superkingdom, phylum, class, order, family, genus, and species. These are typically provided in a spreadsheet separately from the signature collection. The tax module also supports lins taxonomies, for which we have a tutorial.

There are several advantages that this approach affords sourmash. One is that sourmash is not tied closely to a specific taxonomy - you can use either GTDB or NCBI as you wish. Another advantage is that you can create your own custom taxonomic ranks and even use them with private databases of genomes to classify your own metagenomes.

The main disadvantage of sourmash’s approach to taxonomy is that sourmash doesn’t classify individual metagenomic reads to either a genome or a taxon. (Note that we’re not sure this can be done robustly in practice - neither short nor long reads typically contain enough information to uniquely identify a single genome, especially if there are many genomes from the same species present in the database.) If you want to do this, we suggest running sourmash gather first, and then mapping the reads to the matching genomes; then you can determine which read maps to which genome. This is the approach taken by the genome-grist pipeline.

Using sourmash tax to do taxonomic analysis

We recommend using the tax module to do taxonomic classification of genomes (with tax genome) and metagenomes (with tax metagenome). The tax module commands operate downstream of sourmash gather, which builds a minimum set cover of the query against the database - intuitively speaking, this is the shortest possible list of genomes that the query would map to. Then, both tax genome and tax metagenome take the CSV output of sourmash gather and produce taxonomic profiles. (You can read more about minimum set covers in Lightweight compositional analysis of metagenomes with FracMinHash and minimum metagenome covers, Irber et al., 2022.)

The tax metagenome approach was benchmarked in Evaluation of taxonomic classification and profiling methods for long-read shotgun metagenomic sequencing datasets, Portik et al., 2022 and appears to be both very accurate and very sensitive, unless you’re using Nanopore data or other data types that have a high sequencing error rate.

It’s important to note that taxonomy based on multiple k-mers is very, very specific and if you get a match, it’s pretty reliable. On the converse, however, k-mer identification is very brittle with respect to evolutionary divergence, so if you don’t get a match it may only mean that the specific species or genus you’re searching for isn’t in the database.

Using sourmash lca to do taxonomic analysis

The sourmash lca module supports taxonomic classification using single hashes, corresponding to single k-mers, in an approach inspired by Kraken. Briefly, you first build an LCA database using lca index, which takes a taxonomy spreadsheet and a collection of sketches. Then, you can use lca classify to classify single-genome sketches or lca summarize to classify metagenomes.

The lca approach is not published anywhere, but we’re happy to discuss it in more detail; just post to the issue tracker.

While we do not recommend the lca approach for general taxonomic classification purposes (see below!), it remains useful for certain kinds of diagnostic evaluation of sequences, so we are leaving the functionality in sourmash.

sourmash tax vs sourmash lca

Why do we recommend using the tax module over lca? sourmash lca was the first implementation in sourmash, and over the years we’ve found that it is prone to false positives: that is, individual k-mers are very sensitive but are often misassigned to higher taxonomic ranks than they need to be, either because of contamination in the reference database or because the taxonomy is not based directly on genome similarity. Instead of using single k-mers, sourmash gather estimates the best matching genome based on combinations of k-mers, which is much more specific than the LCA approach; only then is a taxonomy assigned using sourmash tax.

The bottom line is that in our experience, sourmash tax is as sensitive as lca, and a lot more specific. Please let us know if you discover differently!

Abundance weighting

By default, sourmash tracks k-mer presence, not their abundance. The proportions and fractions reported also ignore abundance. So, if sourmash gather reports that a genome is 5% of a metagenome, it is reporting Jaccard containment of that genome in the metagenome, and it is ignoring information like the number of reads in the metagenome that come from that genome. Similarly, when sourmash compare compares genome or metagenome signatures, it’s reporting Jaccard similarity without abundance.

However, it is possible to take into account abundance information by computing signatures with -p abund. The abundance information will be used if it’s present in the signature, and it can be ignored with --ignore-abundance in any signature comparison.

There are two ways that abundance weighting can be used. One is in containment queries for metagenomes, e.g. with sourmash gather, and the other is in comparisons of abundance-weighted signatures, e.g. with sourmash search and sourmash compare. Below, we refer to the first as “abundance projection” and the second as “angular similarity”.

Projecting abundances in sourmash gather:

sourmash gather can report approximate abundance information for containment queries against genome databases. This will give you numbers that (approximately) match what you get from counting mapped reads.

If you create your input signatures with -p abund, sourmash gather will use that information to calculate an abundance-weighted result. This will weight each match to a hash value by the multiplicity of the hash value in the query signature. You can turn off this behavior with --ignore-abundance.

For example, if you have a metagenome composed of two equal sized genomes A and B, with A present at 10 times the abundance of B, gather on abundance-weighted signatures will report that approximately 91% of the metagenome is A and approximately 9% is B. (If you use --ignore-abundance, then gather will report approximately 50:50, since the genomes are equal sized.)

You can also get count-like information from the CSV output of sourmash gather; the column median_abund contains the median abundance of the k-mers in the match to the given genome.

Please see Appendix B, below, for some actual numbers and output.

Buyer beware: There are substantial challenges in doing this kind of analysis on real metagenomic samples, relating to genome representation and strain overlap; see this issue for a discussion.

Computing signature similarity with angular similarity.

If signatures that have abundance information are compared with sourmash search or sourmash compare, the default comparison is done with angular similarity. This is a distance metric based on cosine similarity, and it is suitable for use in clustering.

For more information on the value of this kind of comparison for metagenomics, please see the simka paper, Multiple comparative metagenomics using multiset k-mer counting, Benoit et al., 2016.

Implementation note: Angular similarity searches cannot be done on SBT or LCA databases currently; you have to provide lists of signature files to sourmash search and sourmash compare. sourmash will provide a warning if you run sourmash search on an LCA or SBT with an abundance-weighted query, and automatically apply --ignore-abundance.

Estimating ANI from FracMinHash comparisons.

As of v4.4, sourmash can estimate Average Nucleotide Identity (ANI) between two FracMinHash (“scaled”) sketches. sourmash compare can now produce a matrix of ANI values estimated from Jaccard, Containment, or Max Containment by specifying --ani (optionally along with search type, e.g. --containment). sourmash search, sourmash prefetch, and sourmash gather will now output ANI estimates to output CSVs.

Note that while ANI can be estimated from either the Jaccard Index or the Containment Index, ANI from Containment is preferable (more accurate). For sourmash search, sourmash prefetch, and sourmash gather, you can optionally return confidence intervals around containment-derived ANI estimates, which take into account the impact of the scaling factor (via --estimate-ani-ci).

For details on ANI estimation, please see our preprint “Debiasing FracMinHash and deriving confidence intervals for mutation rates across a wide range of evolutionary distances,” here, Hera et al., 2022.

What commands should I use?

It’s not always easy to figure that out, we know! We’re thinking about better tutorials and documentation constantly.

We suggest the following approach:

  • build some signatures and do some searches, to get some basic familiarity with sourmash;

  • explore the available databases;

  • then ask questions via the issue tracker and we will do our best to help you out!

This helps us figure out what people are actually interested in doing, and any help we provide via the issue tracker will eventually be added into the documentation.

Appendix A: how sourmash gather works.

The sourmash gather algorithm works as follows:

  • find the best match in the database, based on containment;

  • subtract that match from the query;

  • repeat.

  • when the number of shared hashes between the remaining query and the best match drops below threshold_bp/scaled (or is zero), break out of the loop.

The output below is the CSV output for a fictional metagenome.

The first column, f_unique_to_query, is the fraction of the database match that is unique with respect to the original query. It will always decrease as you get more matches. The sum of f_unique_to_query across all rows is what is reported in by gather as the fraction of query k-mers hit by the recovered matches (unweighted) and should never be greater than 1!

The second column, f_match_orig, is how much of the match is in the original query. For this fictional metagenome, each match is entirely contained in the original query. This is the number you would get by running sourmash search --containment <match> <metagenome>.

The third column, f_match, is how much of the match is in the remaining query metagenome, after all of the previous matches have been removed.

The fourth column, f_orig_query, is how much of the original query belongs to the match. This is the number you’d get by running sourmash search --containment <metagenome> <match>.

f_unique_to_query      f_match_orig  f_match                f_orig_query
0.3321964529331514     1.0           1.0                    0.3321964529331514
0.13096862210095497    1.0           1.0                    0.13096862210095497
0.11527967257844475    1.0           0.898936170212766      0.12824010914051842
0.10709413369713507    1.0           1.0                    0.10709413369713507
0.10368349249658936    1.0           0.3134020618556701     0.33083219645293316

Where there are overlapping matches (e.g. to multiple E. coli species in a gut metagenome) the overlaps will be represented multiple times in this column.

A few quick notes for the algorithmic folk out there –

  • the key innovation for gather is that it looks for groups of k-mers in the databases, and picks the best group (based on containment). It does not treat k-mers individually.

  • because of this, gather does not saturate as databases grow in size, and in fact should only become more sensitive and specific as we increase database size. (Although of course it may get a lot slower…)

Appendix B: sourmash gather and signatures with abundance information

Below is a discussion of a synthetic set of test cases using three randomly generated (fake) genomes of the same size, with two even read data set abundances of 2x each, and a third read data set with 20x.

Data set prep

First, we make some synthetic data sets:

  • r1.fa with 2x coverage of genome s10

  • r2.fa with 20x coverage of genome s10.

  • r3.fa with 2x coverage of genome s11.

then we make signature s10-s11 with r1 and r3, i.e. 1:1 abundance, and make signature s10x10-s11 with r2 and r3, i.e. 10:1 abundance.

A first experiment: 1:1 abundance.

When we project r1+r3, 1:1 abundance, onto s10, s11, and s12 genomes with gather:

sourmash gather r1+r3 genome-s10.sig genome-s11.sig genome-s12.sig

we get:

overlap     p_query p_match avg_abund
---------   ------- ------- ---------
394.0 kbp     49.6%   78.5%       1.8    genome-s10.fa.gz
376.0 kbp     50.4%   80.0%       1.9    genome-s11.fa.gz

  • approximately 50% of each query matching (first column, p_query)

  • approximately 80% of subject genome’s contents being matched (this is due to the low coverage of 2 used to build queries; p_match)

  • approximately 2.0 abundance (third column, avg_abund)

  • no match to genome s12.

A second experiment: 10:1 abundance.

When we project r2+r3, 10:1 abundance, onto s10, s11, and s12 genomes with gather:

sourmash gather r2+r3 genome-s10.sig genome-s11.sig genome-s12.sig

we get:

overlap     p_query p_match avg_abund
---------   ------- ------- ---------
0.5 Mbp       91.0%  100.0%      14.5    tests/test-data/genome-s10.fa.gz
376.0 kbp      9.0%   80.0%       1.9    tests/test-data/genome-s11.fa.gz

  • approximately 91% of s10 matching

  • approximately 9% of s11 matching

  • approximately 100% of the high coverage genome being matched, with only 80% of the low coverage genome

  • approximately 2.0 abundance (third column, avg_abund) for s11, and (very) approximately 20x abundance for genome s10.

The cause of the poor approximation for genome s10 is unclear; it could be due to low coverage, or small genome size, or something else. More investigation needed.