I'm preparing the pipeline for a new release of the SBDI-GTDB database, based on GTDB R11-RS232. In doing this, I also made sure the pipeline can run phylogenetic placement (using code also available standalone in #C04NLA7QHH9) to the GTDB reference trees optimized for 16S sequences (just speficify `--dada_ref_taxonomy sbdi-gtdb --run_pplace`; phylogenetic placement unfortunately requires quite a lot of memory, ca. 100 GiB). To investigate how phylogenetic placement performs compared to the standard DADA2 methods

assignTaxonomy

and

addSpecies

, see attached html. In short, phylogenetic placement gives a more specific assignment than the sequence composition and sequence identity-based methods in DADA2. Hopefully, this will be available soon in the next stable release of Ampliseq. (PR coming soon, @Daniel Straub.)

That looks indeed great. However, as you have noted, some species annotations were also downgreaded to e.g. genus. Also, its unclear whether the assigned taxonomies are correct, or did you check that? 😅

No, I didn't check that. My feeling is that phyloplacement is better than sequence composition, but that's untested at the moment. A colleague is testing this with a "leave one out" approach. When I know the results, I'll post here

Looks interesting! I’m curious how this compares to methods other that the RDP (QIIME2's feature classifier, SINTAX, etc)?

Yes, that would indeed be interesting. I always use the

assignTaxonomy

results so I have very little experience of the other methods. The same colleague who's doing a more thorough analysis of this has earlier found that sintax works better for some taxonomic groups (eukaryotes) so he might look into this too.

I’ve also had some conversations w/ Ben Callahan about the use of

assignTaxonomy

, but more in relation to long reads. His recommendation was to use local alignment approaches (as was used in his PacBio paper). Possibly in combination with approaches to disambiguate ties or close matches using LCA approaches.

You should try sbdi-gtdb in Ampliseq. We preprocess all 16S sequences from the GTDB through Sativa (https://github.com/amkozlov/sativa) to ensure sequences have a phylogenetic signal that is consistent with their taxonomy.

Could be worse 😉. We have a lot of users who choose old versions of Silva (primarily due to logistics with comparing results to older data).

Yes, to talk people out of using Silva is indeed difficult sometimes

Maybe we should change the default? Because SILVA is now really outdated compared to GTDB? At least that will force ppl to think about choices in case they want SILVA.

I would support that!

Same. The only advantage I’ve seen re: Silva is inclusion of euk 16S, and in that case we only use those data for identifying/filtering purposes. GTDB for everything else now, unless someone reaaally insists (see my comment above 😉 )

Yes, the chloroplast sequences is indeed a problem. We do like you, plus run pr2 to get a taxonomy of the eukaryotes too.