Include RDP classifier in q2-feature-classifier

From a quick search on the forum and the q2-feature-classifier plugin, it doesn’t seem like the RDP classifier is supported by QIIME 2.

Coming from the perspective of a curmudgeony microbiome researcher switching into QIIME 2, this is pretty unfortunate. What’s the reasoning behind not including it? Is it installation issues (would need to download RDP from SourceForge) or something about the fact that you need java to run it?

If there aren’t any particularly hairy technical reasons why this hasn’t been included, I’d be happy to give it a shot. But before diving into an endless rabbit hole, wanted to make sure I’m caught up with whatever discussions y’all had on this topic!


Great question @cduvallet! As another curmudgeony microbiome researcher who has used RDP for years, I can understand your concern. There are a few reasons:

  1. Installation
  2. Java (code maintenance, dependencies)
  3. The sklearn Naive Bayes classifier available in q2-feature-classifier is effectively the same method (and performs very similarly or better according to our benchmarks), except that it is not as fast as RDP classifier. (and we are making additional improvements to the classify-sklearn method that make it even better)

You are more than welcome to make a q2-rdp plugin!

Thanks for asking!


Cool, didn’t realize there was a paper benchmarking them (I should read the docs more closely…)! Will check this out and almost certainly be satisfied.

Hah, well, those issues #1 and 2 seem hard so pretty unlikely that I’ll go reinvent the wheel here :wink:

Thanks for the response!


conda-forge has an openjdk build, it is probably possible…

1 Like

Hello Claire, I was having exact same question and found your post.

In case you tried the sklearn Naive Bayes classifier available in q2-feature-classifier, would it be possible for you to share the parameters you used ?

classifier documentation:

Thanks a lot


Welcome to the forum @Francis29029!
Chances are Claire used the default settings (confidence=0.7). The performance of the default classifier is quite similar to RDP with identical confidence settings, see this note from above for more details:

1 Like

Yep, I just used default settings! Didn’t really do much/any due diligence on the defaults though, but mostly cuz I assumed @Nicholas_Bokulich and team already did it for me! :slight_smile: