Which taxonomy data type to use for abundance graphs and statistical analyses?

Question

At our institution we have had some debate about this topic. So let me give some background:

Qiime2 produces a OTU table which shows the number of sequences that matched to a specific OTU. It also produces an Taxonomy table which links all the OTUs to their corresponding taxonomic identity.

So what we were discussing was whether to use the number of sequence hits per OTU for alpha and beta diversity, as well as abundance graphs. Or should we use the Taxonomy table and count the number of occurrences for each taxa.

An example of this would be something like the following:

Here are two OTUs:

D_0__Bacteria;D_1__Cyanobacteria;D_2__Oxyphotobacteria;D_3__Chloroplast;D_4__Virgulinella fragilis;D_5__Virgulinella fragilis;D_6__Virgulinella fragilis and D_0__Bacteria;D_1__Cyanobacteria;D_2__Oxyphotobacteria;D_3__Nostocales;D_4__Oscillatoriaceae;D_5__Oscillatoria PCC-10802;Ambiguous_taxa

If we used the first approach then there would be, for example, 3000 sequence matches for the first OTU and 1500 for the second OTU. However, they represent two different species from the same Class. So abundance wise, the first OTU could be represented twice as much as the second. With the second approach they would share the same abundance percentage.

What is the correct approach for these types of analyses according to the bioinformatics community?

Answer 1

The fundamental answer to this question is the role of allele frequency, i.e. population genetics, within taxonomy. The answer of course is that population genetics and population dynamics should be a integral part of taxonomic designation PROVIDING it is not subject to unreported sample bias. Thus it has to be option 1.

However, your definitions I'm not so sure about because the counts per OTU using your literal definition means every SNP within a species. I don't believe that is what you mean, but that is what you have said. Obviously at a taxonomic focus of a single SNP this isn't meaningful because there would be as many OTUs as there are sequences and the species definition would be meaningless.

Both approaches suffer from lacking a clear definition to make a species assignment and this is controversial. If you are using a tree to assign species membership, e.g. 16S at 5% identity, then that would generally be acceptable, but some would question it.

You can of course use deep learning algorithms to do this, but this is only as accurate your training and at present (check Bioinformatics journal) the accuracy is at best 90%, which is very low by comparison to phylogeny.

Thus I think it is a non-question, because allelic diversity is extremely important in vast swathes of biological analysis. Taxonomy is not a disapline in its own right, but a foundational step biologists use to understand the ecotope, ecoregion etc... The only justification for pursing approach 2 is if there is persistent concern in the bias of the sampling process, to the point where the alleleic (species) frequency becomes meaningless.