The Critical Assessment of Metagenome Interpretation and why I’m not a fan

Written by: C. Titus Brown

Primary Source: Living in an Ivory Basement

If you’re into metagenomics, you may have heard of CAMI, the Critical Assessment of Metagenome Interpretation. I’ve spoken to several people about it in varying amounts of detail, and it seems like the CAMI group is working to generate some new shotgun metagenome data sets and will then encourage tool developers to bang on them. (You can also read a short Methagora blog on CAMI.)

I’ve been asked by about a dozen people now what I think of it, so I’m blogging about it now rather than answering people individually :).

tl; dr? I’m not a fan.

First, what’s this subscription nonsense? (We’ll keep you in the loop if you register here.) There are a plethora of ways to keep people in the loop without asking them to subscribe to anything (blogs, Twitter, Web sites, e-mail lists…). Is there a reason to keep the details of the contest secret or hidden behind a subscriber wall? (Answer: no.)

Second, I am told that the contest will be run once, with software that is blind to the content of the metagenome. I understand the idea of blind software evaluation, but this is not a long-term sustainable approach; we need to generate a nice set of orthogonal data sets and then LART people who fine-tune their software against one or another.

Third, it looks like the CAMI folk will make the same mistake as the Assemblathon 2 folk, and not require that the analyses be completely replicable. So in the end there will be a massive expenditure of effort that results in a paper, which will then be a nice static record of how things were back in 2014. Given the pace of tool and technology change, this will have a very short shelf-life (although no doubt tool developers will cite it for years to come, to prove that IDBA was once worse than their own assembler is now). Why not re-run it every 6 months with the latest versions of softwares X, Y, and Z? We have plenty of ways to automate analyses, and there is simply no excuse for not doing so at this point. (ht to Aaron Darling for alerting me to

Fourth, there are several mock metagenome data sets that are already underanalyzed. For example, we’re currently working with the Shakya et al. (2013) data set, but I don’t think anyone else is (and it’s pretty clear most people don’t realize what a stinging indictment of 16s this paper is – or, at least, don’t care. Discussion for another time.).

So why are we not poking at existing data first? I don’t know, but I suspect it speaks to an underlying bias against re-analyzing published data sets, which we really need to get over in this field.

As I said during the Q&A for my BOSC 2014 keynote, I’m not a big fan of how we do benchmarks in bioinformatics; I think this is a fine exemplar.

I probably won’t participate in the CAMI benchmark, not just for the above reasons but because I simply don’t have the people to do so at the moment. …plus, as with Assemblathon 2, they’re going to need reviewers, right? ;)

Note that Aaron Darling raised many similar points at ISME, and they were apparently very well received. Maybe CAMI will adapt their approach in response, which would be great!


p.s. Thanks to Alice McHardy and Aaron Darling for their detailed discussions of this with me!

The following two tabs change content below.
C. Titus Brown
C. Titus Brown is an assistant professor in the Department of Computer Science and Engineering and the Department of Microbiology and Molecular Genetics. He earned his PhD ('06) in developmental molecular biology from the California Institute of Technology. Brown is director of the laboratory for Genomics, Evolution, and Development (GED) at Michigan State University. He is a member of the Python Software Foundation and an active contributor to the open source software community. His research interests include computational biology, bioinformatics, open source software development, and software engineering.