ctb
commented
4 months ago Hi @ezherman I have nothing but intuition to offer, I'm afraid! Hopefully some others will step up with some better experience!
I am in the process of determining whether sourmash is the right tool for my metagenomic analyses. I am mostly using ONT data at the moment. Similar to previous posts (e.g. #2236), I have found that sourmash cannot classify most of the kmers.
Right! I would expect that!
My question is, do you need most k-mers to be classified, or would you be satisfied with having enough k-mers be classifiable to get accurate results?
There are two, mmm, schools of thought on taxonomy. One is taxonomic binning, where you are trying to assign each sequence to a specific taxonomic group. The other is taxonomic profiling, where you are trying to identify taxonomic groups present in your entire data set.
sourmash has (so far) mostly been used for taxonomic profiling, where it is very accurate for reasons I think we can explain (combinations of k-mers, basically).
I would not suggest using sourmash directly for binning (yet), for a few reasons - the main one being that we are only just now getting to the point where it performs well with the lower scaled value needed. (See the branchwater plugin for our rather successful efforts to improve the performance of gather).
However, if you want to get binning-like results, there's a third option here, and it's one you might consider: sourmash actually does genomic profiling of the composition of a metagenome with sourmash gather, and then "promotes" that to a taxonomy with sourmash tax. You could generate a genomic profile of your entire metagenome against your Very Large Database, then map your ONT reads directly to the resulting genomes. Depending on what you want, this might get you there - e.g. if you wanted to find the relevant genomes (and hence taxonomies) for your data set, you will find them this way.
We have a workflow that implements this for Illumina, called genome-grist. Happy to chat more about that here or there.
I was wondering whether there is a "best practice" for using sourmash with ONT data. To aid the discussion, I thought I'd list all the potential processes and parameters that I think may help. Before trying to optimise them all, I thought I'd check whether others have had any success!
(thank you for this summary!)
- The effect of the average read Q score. Given comments that suggest the low classification rate is due to sequencing errors (e.g. Improving results for nanopore #2236 (comment)), I have been considering filtering for, say, an average read Q score of 20. However, this can massively reduce the number of reads.
Agreed, this is a bad idea!
- Effect of k-mer size. The main sourmash guide suggests using k=31 or k=51, with the latter recommended for strain-level resolution (here). I reckon this recommendation came out of Illumina data. In contrast, one suggestion was to choose k as small as possible given the average genome size for an optimal sensitivity/specificity trade-off (adjust default thresholding for prefetch and gather? #2360 (comment)). For a 6Mb genome, this would come to k=11 or k=13.
I was going to say k=15 ;). I agree. The only challenge here is you'd need to build your own database at the k-mer size(s) of interest. That has become much faster and easier with manysketch from the branchwater plugin, but it's still an extra step.
- Effect of k-mer abundance. Sourmash sig filter allows for removal of kmers of a certain abundance (e.g. 1), which are more likely to have come from sequencing errors.
Yep. It's likely to reduce your computational costs and maybe your sensitivity, but should not have any effect on specificity.
- Effect of
--threshold-bp: computational requirements go up with a lower threshold, however I understand this to also increase sensitivity (adjust default thresholding for prefetch and gather? #2360 (comment)).
Yep. I would say that the computational requirements are less of an issue now, with the branchwater plugin.
If anyone has had success with these adjustments on ONT data, or is aware of any other modifications, I would really appreciate your input.
If you can identify a simple benchmark data set for us to try out (Maybe the one from Portik et al would work - @bluegenes ?) we could potentially do some of this for you...
ezherman
I am in the process of determining whether sourmash is the right tool for my metagenomic analyses. I am mostly using ONT data at the moment. Similar to previous posts (e.g. https://github.com/sourmash-bio/sourmash/issues/2236), I have found that sourmash cannot classify most of the kmers.
I was wondering whether there is a "best practice" for using sourmash with ONT data. To aid the discussion, I thought I'd list all the potential processes and parameters that I think may help. Before trying to optimise them all, I thought I'd check whether others have had any success!
The following seem relevant:
--threshold-bp: computational requirements go up with a lower threshold, however I understand this to also increase sensitivity (https://github.com/sourmash-bio/sourmash/issues/2360#issue-1446498334).If anyone has had success with these adjustments on ONT data, or is aware of any other modifications, I would really appreciate your input.