DADA2 versus multiple.split.libraries.py

Dear all, I have read that one should expect fewer features when applying DADA2 compared to the number of OTUs that would be identified and accepted after multiple.split.libraries.py in QIIME1.

I have tested this on a subpopulation supplying me 106 samples (sputum), and I now have some questions regarding the outputs and methods. First in Q1 I applied vsearch for chimera removal and end up with 9900 OTUs (97%similarity). DADA2 leaves me with 3500 features. Now in Q1 I did not remove the primer regions as I ran pick_open_reference_otus.py. In DADA2 the primer regions are removed.

Still I find it to be overwhelming amounts of both OTUs and features so in Q1 I applied the filtering as advised by Boculich - and 9900 OTUs are redused to 370 :open_mouth: In Q2 - even though not recommended I decided to filter out all features found in only one single sample, and also filtering out features represented by less than 5 sequences. Then I end up with 1000 features.

I still think this is worringly many features and now I wonder if there is other filtering measures I have ignored, or if this simply is what I should expect from my fine group of chronic lung patients providing me with sputum :grinning: And could the Q1 OTU load be explained simply by me ignoring the primer regions?
Solveig

1 Like

Hi @stangedal,
This is a really great question!

The short answer: I would trust dada2 results over the QIIME1 results with abundance filtering. You could fiddle with the dada2 parameters a bit to make them a little more stringent (and @benjjneb may offer some advice in that regard). There may still be other issues inflating the feature count (e.g., contamination or PCR errors that aren’t being caught by dada2) but we don’t currently have methods that can identify such issues. This may be the best you can do at the moment. Science is a work in progress.

Now the long answer(s) (and questions):

Is this necessarily unexpected? This does sound high but I am unfamiliar with this sample type — what do others report for sputum samples? What’s the average feature count per sample after you rarefy at an appropriate depth? (subjects may display a high degree of inter-individual variation and hence have many unique features that are not shared across the entire group of patients; average features/sample could be much lower, though)

The filtering described in that paper was quite stringent, to overcompensate for the fact that QIIME1 OTUs are very erroneous. Sounds like you actually followed the instructions in that paper (:astonished: :1st_place_medal: :fireworks: :clap: ). If you are not interested in entertaining the possibility that dada2 may enable exploration of the rare features detected in these samples, you could follow a similar approach here — but I’d still recommend dada2 over OTU picking.

Yes, at least partially. The presence of primers will affect the OTU clustering — primer sites will be very highly conserved across all seqs in your data, so the similarity between seqs with primers present would be higher than the same seqs with primers trimmed, effectively reducing your number of OTUs since seqs that are borderline 97% similar will suddenly pass that threshold.

Without an internal standard (e.g., a mock community) to help calibrate, I would not overthink this nor try tinkering too much unless if these values just seem wildly off to you. But please do share if you find an approach that delivers more palatable results.

I hope that helps!

3 Likes

Thank you so much for your answer and advise. It was a valuable exercise investigating the means of features by sample. It made me a lot more confident about the results.

As of to day it is hard to say what others have found in sputum - a quick search could not give any comparable studies where DADA2 is reported to be used. But I am sure DADA2 is implemented in different research environments on sputum too, so I believe we will be getting more information within months, or at least a few years…

After discussing it in our research group we end up thinking that the biological value of a feature found in 1 patient, and built by 1-4 sequences is hard to believe in, so we will filter at that level our first run through the data, and then repeat the analyses on the same samples - without additional filtering after DADA2. And we also plan to filter even more and see what that reveals, but we have to discuss that a little further.

Thanks again for valuable input!
Solveig

2 Likes

Thanks @stangedal!

That seems quite reasonable to me, and I think that’s the bottom line — you need to go with what you believe in and what is most biologically relevant. Features observed in one patient/sample based on 1 sequence (provided the sample has adequate sequence coverage) probably isn’t biologically important, even if it is a true observation.

This topic was automatically closed 31 days after the last reply. New replies are no longer allowed.