Qiime2 outperforming Mothur?


Hi, I was talked now QiimeII is much better than other packages, and I was given this technical note to read doi: 10.1093/gigascience/giy054

What is your opinion on this? if we work with environmental samples, like soils or sediments, should we switch from Mothur to QiimeII or is there a way to get with Mothur the same performance this people had with QiimeII, maybe including some features in Mothur?

Thank you!


See my comments on the other thread.


Thank you Pat! but the only reason they had more “reliable amount of OTUs recalled” in this publication when they performed the analysis on the same datasets obtained from mock communities is because the singletons were removed? So, you do not agree with singletons being likely PCR/sequencing artifacts?
I still don’t get the apparently so big differences between Mothur/Qiime and the new QiimeII (which they propose as the best option).


No, I don’t remove singletons. Take a single community that you get 5000 and 15000 reads from in separate replicates. Do you automatically reject the singletons from the 5000 read replicate? Do you then remove the tripletons from the 15000 read replicate? Both decisions will influence how you measure the community’s diversity and it’s similarity to other communities. FWIW, this is also why we leave the rare sequences and rarefy.

Sure there is garbage down in those rare classes, but there is also garbage that shows up many times because a PCR error or chimera will be compounded over multiple cycles. Leave all the data there.


Ok! by rarefy here do you mean to cut at the same amount of sequences per sample?
And what is your opinion on why they found such big differences working with soil samples when they used qiimeII, qiimeI and Mothur? what do they mean exactely with better “recall” and “precission”?
Sorry I don’t fully understand that :frowning:


Rarefaction is drawing the same number of reads from every sample and calculating your alpha/beta diversity metric. Then you put the reads back and do it again. You do it 1000 times and take the average. That’s rarefaction.

If I had to guess, mothur uses a more conservative confidence threshold than QIIME. From the paper it makes it clear that they both use the naive Bayesian classifier. We use 80% as a threshold, I suspect they use 50%. It’s not a comparison of mothur vs QIIIME, it’s a comparison of thresholds. A better test would be to do a leave one out test where they take the full database and pull out a sequence and classify it against a database missing that sequence. Then repeat for all of the sequences in the database.

For the jargon please see


Thank you Pat! I’ll look into that link. It’s only that there was such big push in this course I attended on using the new QiimeII instead of any other package (including the old Qiime) and there was this technical note that I wanted to know the reason why, and be convinced and not go and switch to the new QiimeII only because “they sayd that”.


Who taught the course and where was it?


It was in Argentina, and one teaching assistant (as far as I understood) is doing a postdoc abroad (guess Southafrica) and they got a paper rejected and were asked to use QiimeII instead of QiimeI. Then I looked for information supporting the idea they transmitted (now it’s QiimeII and no QiimeI or mothur or any other) and then I found that technical note. And decided to ask here for experienced opinion and explanation. I used Mothur only twice and barely could learn it, but I felt satisfied and find it easier to use than Qiime (except for few features like having a tree for unifrac distances but using the otu approach).