kmers for taxonomy-free beta-diversity measures: an opportunity to integrate amplicon and shotgun analysis?

General Discussion
1

Greetings Qiime-izens, :qiime2::popcorn:

TL;DR below. :stopwatch:

I have some environmental megagenomic samples that have been analyzed with 16S and shotgun sequencing. So, we ideally want our beta diversity measures to be on the same footing between our two different datasets (amplicon vs shotgun).

One way to do this is taxonomy-based, using the GreenGenes2 database. The handful of publications I've found using this approach are all based on human gut microbiome samples; the human gut is a highly sequenced environment well-represented in reference databases. We too are planning on using taxonomy-based dataset integration, such as described in Usyk et al 2023.

One thing that bothers me about this approach is the discrepancy in how much of each dataset is classifiable: 70-90+% of ASVs will provide some taxonomic info, whereas maybe 5% of shotgun reads will be classifiable, with assembly boosting the classifiable shotgun % somewhat. It seems like the difference in % classification between the two datasets could easily overwhelm the beta-diversity signal.

On the other hand, the q2 kmerizer paper shows that diversity tests can be performed accurately on kmers of amplicons, independently of classification. :flexed_biceps:

Similarly, in Sourmash for example, taxonomy-free difference measures (cosine / angular dissimilarity) of shotgun data can be measured with kmers on unclassified unassembled reads.

So, why not use taxonomy-free kmers as a way to standardize beta-diversity measures between amplicon and shotgun data? :thinking:

What i like about this approach is that the beta diversity measures for both amplicon and shotgun datasets can be based on all of the respective samples, rather than just the classifiable fraction.

I also like that it does not rely on only marker genes or reference databases: for our environmental sequences, it will capture the effects of novel and referenced taxa equally well.

I understand that there are still significant apples-to-oranges issues when integrating amplicon and shotgun data, for example differences in sequencing depth and alpha-rarefaction, as well as PCR bias.

I also understand that we can use a combination of different tests to improve our overall picture of the data. So, I can use both kmer and taxonomy based analysis, which will give me different but related info.

TL;DR:

  1. I want use kmer-based taxonomy-free diversity measures to integrate analysis of amplicon and shotgun data. I think that conclusions about whether Group A and Group B are different is more directly comparable between amplicon and shotgun data with this method.
    I'm second-guessing myself because I have not encountered anyone doing this in the literature.

  2. I'm also second-guessing my understanding of the shortcomings of using taxonomy-based beta-diversity measures between amplicon and shotgun data. Wouldn't the effect of the large difference in % classification between amplicon and shotgun data often be much larger, and impossible to distinguish, from the beta diversity signal?

Certainly I need to continue to read up on this topic. I'd be grateful for any comments or suggestions from the forum! :folded_hands:

2 Likes
2

Hi @sibilant ,
It is a very cool idea! I am sceptical that it would work, but it is worth a shot if you have paired samples you could do something like sourmash + kmerizer ordinations on those paired samples and then generate a procrustes plot to see how well these correlate.

Why am I sceptical? The shotgun metagenome data would cover a much larger sequence space, with much larger kmer diversity. I am also not sure if kmerizer and sourmash would be directly comparable, as if I recall correctly sourmash is generating a MASH index from the kmer profiles (kmerizer is not).

On the other hand, you could extract 16S reads from the shotgun data, e.g., with sortmerna, so that you are looking at kmers from the same region. If this is a meta-analysis and you have different 16S subregions in the amplicon data, you could go even further by mapping all reads (shotgun and 16S) to full-length 16S reference seqs. Then you could pass it all through kmerizer (though at that point by mapping to reference seqs you will also have putative taxonomy, reference trees, but also lose reads that do not map so maybe this is more stringent than your plan)

1 Like
3

Thank you so much Dr. Bokulich! :light_bulb:

The shotgun metagenome data would cover a much larger sequence space, with much larger kmer diversity.

So, with the kmer approach I was thinking, the signal for gene diversity (16S only amplicons, vs all different kinds of shotgun genes) would overwhelm any difference in community members. Whereas with a taxonomy-based approach, the unit being measured would be taxa, rather than genes. That makes a lot more sense. :sweat_smile:

Extracting 16S reads from shotgun data is an interesting option too: might be a way to compare the alpha-rarefaction between shotgun and amplicon data; maybe mediate the effect of PCR bias for certain taxa under-represented in the amplicon data.

Thanks again very much! :seedling: :orange_heart:

1 Like