Qiime Vs Mothur : Why Use One Over The Other?

Question

[amcrisan]

In my own "metagenomic" (amplicon based) analysis, I've had a preference for using Mothur. My reasons for it are as follows:

1. Seems like it does a good job (based on analysis with Mock community)
2. It has a simple SOP that was straight forward to follow and is published

I am very new to community analysis, so Mothur seemed like an easy first step (and I've had good support on the forum..so extra bonus). A long the way I've gotten questions about using QIIME. After looking into the differences between the two, I still can't fully grasp what the major differences between the methods are to motivate one person to prefer one method compared to the other. Does one method make the other obsolete?

I would really appreciate if someone who has worked with either (or both) methods could shed some light on why pick one over the other.

[Cliff Beall]

Just to muddy the waters, there is uParse as well: http://www.nature.com/nmeth/journal/v10/n10/full/nmeth.2604.html

[Josh Herr]

uparse (like uclust, cdhit, etc) is a clustering algorithm and not really a "stand alone" analysis program/pipeline like mothur and QIIME. In fact, the are lots of options for clustering from within QIIME and mothur, and both can run uparse from within.

[Istvan Albert]

Interesting, we're currently running a few large classifications, so we'll take a close look. Thanks for the link.

Answer 1

[Istvan Albert]

There is a fundamental almost philosophical difference in how the tools are developed

Mothur is a single program that re-implements a large number of very useful algorithms into a single, high performance standalone executable program for each platform: linux, mac and windows . In my opinion it is one of the most amazing feats of bioformatics software engineering especially considering that is being developed by only two people.

QIIME is a python interface (glue) that connects a very large number of disparate programs and what QIIME really does is transforms the input/outputs of these and allows you to feed one program into the other. You will need to install a very large number of dependent programs to use QIIME. In fact it is not even correct to say that you used QIIME to compute something, you are using the programs that QIIME runs for you: pynast, uclust etc.

Now because in QIIME each program is developed independently some by entire groups some of these may offer higher performance than what is implemented in mothur. On the other hand the unified interfaces of mothur make it far more consistent and better documented.

Answer 2

[pschloss]

Istvan gets it pretty right. But I would disagree with his statement that "some of these may offer higher performance than what is implemented in mothur". I don't actually know of any examples of this. Some examples of where we've improved things

• NAST - was originally closed source, we opened it up (align.seqs) and it is light years faster than what the pynast paper claims. Furthermore, greengenes actually changed their algorithm to something that was worse, we optimized it, made it parallel, and opened it up to be usable with other databases
• PyroNoise - Quince's interface is horrible and requires the user know bash and perl (um, I don't really know bash). We streamlined that.
• Classifier - The RDP's classifier is built in Java and is very slow. We made it faster by translating it and parallelizing it.
• UChime - this is an example of where there was source code that we just ported directly into mothur with little to no modifications
• There are also a few packages that we effectively maintain at this point because the original developers are MIA

As for QIIME...

• I'm fairly certain that QIIME is much better than mothur at visualization, but ... you can always just use R (and those horrible black backgrounded ordinations, uck!)
• They seem devoted to the greengenes alignment, which is just dreadful. I suppose you could say we're tied to the RDP taxonomy, but that's just because I'm too lazy to change the SOP to the greengenes database, and I think bacterial names are largely like the points on "Who's Line is it Anyway". Our wiki has the gg databases.
• They seem to be moving towards a database dependent approach that depends on a close source software package (USEARCH) that when we benchmarked for a non-db approach did worse than average neighbor. I think this is because they are dealing with a lot of crappy data from the global survey project they have and can't do a true OTU-based approach because the computers will explode

I've not interacted much at all with Rob or the other QIIME developers, but it's really hard to believe that they're anywhere as helpful as Sarah Westcott, who is just awesome. There are users that just drive me nuts and she can help them like they were her 7 yr old and she makes them the center of her world. So why does any of this matter to you - Ms/Mr. User? I would like to think that our improvements make the interface and execution better, that our SOPs and experimentation with a focus on error rate reduction makes the analysis better, that people appreciate our responsiveness, documentation, and attempts to reach out to people. I'd love to hear from QIIME users about what they prefer about QIIME over mothur and what we might be able to do to get their support.

Thanks for bringing this up and asking my opinion!

[Istvan Albert]

Thanks for chiming in - it is always great to get input from the creators and developers of the various tools and techniques

[Daniel]

*Qiiming in (Sorry, I had to...)

[Istvan Albert]

I agree! that's really funny! :-)

Answer 3

[Skeletor]

I've played around with both myself, and would like to give my $0.02.

• I have had pretty different results using both tools even though they both implement a lot of the same algorithms. Without running them on a mock dataset, it's tough to say which is better.
• Pat Schloss is very active on his forum, and has been very helpful. I've never contacted the Qiime people, so I can't comment on their community.
• The mothur approach seems a bit more rigorous to me. I'm assuming that most users will base their pipeline off of the tutorials for each site. Qiime doesn't include chimera checking in theirs - http://qiime.org/tutorials/tutorial.html, although they do include tools for it
• If you don't like running virtual boxes, Qiime you can run into dependency hell for some Linux distros.
• I like hearing people try to pronounce Qiime

Answer 4

[alk224]

I was largely reintroduced to the linux world and bioinformatics through the use of qiime. I have found it very helpful in learning how to do command-line operations and to gain confidence in my rather coarse bioinformatic skills. I have some data that I keep meaning to process via each pipeline, but just haven't gotten around to it yet. At conferences I have seen fantastic visualizations produced (at least from the data) through each package (qiime,mothur,uparse), so my sense is that they all perform the desired functions, though I agree it is debatable which really is best. One is slower here, one is faster there, and often there is a trade-off between speed and accuracy.

The qiime forum is very active and users generally get their questions answered rather quickly. Some of the moderators are extremely quick to identify a users problem and help them get to resolution.

I do like the ability to only run the scripts you want/need for a given dataset. Such freedom can be wielded very well or very poorly, depending on the user (eg, not bothering to do chimera checking).

I ran qiime a few years ago in a virtual box, which was a super lame experience since my laptop was already old and slow. Then I installed it natively on a desktop in lab, which was also a super lame experience because it took me so long to get all the dependencies installed and in the right places. Now we have a linux-based dual xeon workstation (24 cores) with a bunch of ram and disk space and I maintain the qiime distribution with qiime-deploy tool. Installation and upgrading couldn't be easier, and there is no point trying to run qiime on a wimpy laptop if you intend to process a miseq run in an afternoon. We also recently gained access to a cluster and a sysadmin takes care of all that baloney for us.

You can also submit requests to the qiime team to add functionality. In fact, after asking a question about filtering OTUs from a raw OTU table, one of the coders wrote me a custom script to try to improve my filtering based on my concerns with existing methods. When it didn't work quite right, he fixed it within a day. Still evaluating its utility, but I think it shows the strong support of the qiime community.

When it comes down to it, I think the Knight and Schloss labs both really just want to be able to do sound science. I think parallel development of the different platforms is essential so that we don't get stuck in one paradigm when someone else can see a flaw and provide an effective alternative.

Answer 5

[shizhaojingszj]

Now conda/bioconda is the default tool to install QIIME.