While I was on the road last week–ironically to do workshops including one on ENCODE data in the UCSC Genome Browser, a conflama erupted over a new paper that was published essentially spanking the ENCODE team for some of the claims they made. Some of the first notes I saw:
— Daniel MacArthur (@dgmacarthur) February 21, 2013
— Stephen Henderson (@Malarky67) February 21, 2013
— Gene Fault (@GeneFault) February 21, 2013
Luckily I happened to be in a library when this drama broke, so I ran to an unoccupied workshop computer while Trey talked about the Table Browser and read the paper quickly. I will re-read it when I have more time, but I wanted to offer my initial thoughts while the chatter is ongoing.
Subsequently I saw a range of reactions to this: PZ Myers says ENCODE gets a public reaming; Mick Watson’s Dear ENCODE…. ; Homolog.us’ critique of Mick Watson’s response Biomickwatson’s Ridiculous Criticism of ENCODE Critics ; Biostars’ forum ENCODE commentary from Dan Graur …and more. I’m sure there will be further fallout.
My first thoughts were that the paper was the snarkiest scientific paper I have ever read, and I thought it was hilarious. I also think some of the criticisms were completely valid. Some less so.
First I should establish some cred on this topic, and explain my role. I was not part of the official ENCODE analysis team, and was not an author on any of the papers. As OpenHelix we were engaged in some outreach for the UCSC Genome Browser Data Coordination Center–but we worked for and reported to them and not to the main ENCODE group. As such, we delivered training materials and workshops for several years, and although these touched on various data sets presented by many ENCODE teams, we did not have contact with other teams. The materials were aimed at how to locate and use the ENCODE data in the UCSC framework. (ENCODE Foundations and ENCODE Data at UCSC were the recorded materials). However, we are now no longer receiving any ENCODE-related funds in any manner.
So I was exploring ENCODE data a lot earlier than most people. I was making discoveries and finding out interesting new things years ago. And I was also with new users of ENCODE data in workshops around the country. This is the framework that you should use to assess my comments.
On the immortality of television sets….
In the Graur et al paper, there are a number of aspects of the ENCODE project that come under fire. The largest portion of this was aimed at the claim of 80% functionality of the genome. This statement caused problems from day 1, and I agree that it was not a well-crafted statement. It was bound to catch media attention and it irked pretty much everyone. Nearly immediately Ewan Birney tried to explain the position, but most people still found this 80% thing unsatisfying and unhelpful. And I think the Graur et al paper presents why it was so problematic pretty clearly.
Another criticism of the work is that the ENCODE project was focused on cell lines.
“We note that ENCODE used almost exclusively pluripotent stem cells and cancer cells, which are known as transcriptionally permissive environments.”
I understand this concern and even raised it myself in the past in the workshops. But there are 2 important things to note about that: in order to get everyone sufficient sample material to enable comparisons across techniques, technologies, and replications, it would not be possible to use human tissue samples. It just would be physically impossible. Further, a lot of non-ENCODE experimental work is carried out in these cell lines and understanding the difference among cell lines may be incredibly useful in the long run. Making better choices about which ones mirror human conditions, or not using the wrong cell line to test some compound if it’s missing some key receptor could be great information to have. I wish there had been one of the papers that characterized the cell lines, actually.
But another thing everyone missed: STEM CELLS. We now have the largest set of genome-wide data on human embryonic stem cells. This has been information that was particularly hard to obtain in the US, but now everyone can look around at that. I was really sorry to see that aspect of this project got no love whatsoever.
But besides that, the mouse ENCODE project did deliver tissue data. But we can share mouse strains and treatment protocols to get sufficient materials. Additionally the modENCODE project got some really fascinating information on developmental stages that we couldn’t get on humans. I think all of these features are missing in the snark-fest.
Another criticism in the paper is the sensitivity vs specificity choice for reporting on the data.
At this point, we must ask ourselves, what is the aim of ENCODE: Is it to identify every possible functional element at the expense of increasing the number of elements that are falsely identified as functional? Or is it to create a list of functional elements that is as free of false positives as possible. If the former, then sensitivity should be favored over selectivity; if the latter then selectivity should be favored over sensitivity. ENCODE chose to bias its results by excessively favoring sensitivity over specificity. In fact, they could have saved millions of dollars and many thousands of research hours by ignoring selectivity altogether, and proclaiming a priori that 100% of the genome is functional. Not one functional element would have been missed by using this procedure.
Maybe the Graur et al team thinks that it should have been the other way. That’s fine–they can take all of this data and re-examine it, reprocess it, and deliver it with their thresholds. But I think at this time over-prediction is not the worst sin. Some of this technology is still being worked on. Some of techniques will undoubtedly be refined as we go forward. But some of that will shake out once we look at regions and understand why some calls should or shouldn’t be made. Certainly there are going to be artifacts. But there may also be subtle and useful things that researchers on a specific topic and with interests in a specific region will be able to suss out because they had some leads. Maybe some won’t pan out. But certainly that’s not impossible with under-prediction or false negatives either.
I don’t know how many of you have stood in front of rooms of researchers and opened up new data sets to them. I’ve done this quite a bit. I have heard the giggles of a researcher at NIH who was delighted to discover in our workshop that GATA1 binding evidence was present in front of a region she was interested in–and this evidence looked very solid to me. This data came from ENCODE years ago, and she could go back to her lab that afternoon and start to ask new questions and design new experiments long before the controversial statements. Just the other day there was a researcher who found new RNA-seq signals in an area he cares about. Will these turn out to be something? I don’t know. But he was eager to go back to the lab and look harder with the new knowledge.
Big science vs. small science
Another segment of the Graur paper is called “Big Science,” “small science,” and ENCODE”. I tell researchers in workshops that they need to take the leads they get from this and look at it again, confirm it, and poke around with other tools and other cell lines or tissues. But I have seen that the ENCODE data has offered new paths and new ideas to researchers. As I wrote a while ago, ENCODE enables smaller science and people who had no contact with the initial project are making new discoveries with this data. And I think this statement is unfair:
Unfortunately, the ENCODE data are neither easily accessible nor very useful—without ENCODE, researchers would have had to examine 3.5 billion nucleotides in search of function, with ENCODE, they would have to sift through 2.7 billion nucleotides.
Most researchers don’t need 3.5 billion or 2.7 billion nucleotides. But they are very interested in some specific regions, and many of those regions now have new and actionable information that these researchers didn’t have before. And it’s not hard to access this–although we would love to have been funded to do more workshops to show people how they can get to it*.
So in short, I thought the spanking was funny and partially deserved. Some of it was unwarranted. I was a bit surprised to see this level of snarkiness in a scientific paper rather than a blog post or some other format, and I think if that became a publishing trend it might not serve us well. But we are also coming to a point where the literature is less important than the data–because the data isn’t in the papers anymore. What will matter is what we see downstream as people use the ENCODE data. And I hope they do, because I think there’s gold in there. I’ve seen some. But you’ll have to verify it. I think the saddest thing would be is if the drama on the claims made at the end cause people to walk away from the good that came from this. That would be a huge waste.
*If anyone asked me (not that anyone has), I think that outreach on big data projects should be improved in a number of ways. There should be a branch of the project whose only role is outreach–not attached to a specific project team–that has access to all of the teams, but can still maintain some distance. It would help to understand what new users face when they see the project. Often we find that teams on software or data projects are a bit too close to the materials and need to understand what it’s like to be an outsider looking in. And we find that people tell us things that they might not be willing to say to the development team directly, which can be very useful feedback. This is not specific to ENCODE but I have seen this numerous times in other projects as well.
Graur, D., Zheng, Y., Price, N., Azevedo, R., Zufall, R., & Elhaik, E. (2013). On the immortality of television sets: “function” in the human genome according to the evolution-free gospel of ENCODE Genome Biology and Evolution DOI: 10.1093/gbe/evt028
Rosenbloom, K., Sloan, C., Malladi, V., Dreszer, T., Learned, K., Kirkup, V., Wong, M., Maddren, M., Fang, R., Heitner, S., Lee, B., Barber, G., Harte, R., Diekhans, M., Long, J., Wilder, S., Zweig, A., Karolchik, D., Kuhn, R., Haussler, D., & Kent, W. (2012). ENCODE Data in the UCSC Genome Browser: year 5 update Nucleic Acids Research, 41 (D1) DOI: 10.1093/nar/gks1172
Update to add more blowback:
Josh Witten at The Finch and Pea: So I take it you aren’t happy with ENCODE…
John Farrell at Forbes: ENCODE Papers Get A Fisking
Jalees Rehman at SciLogs: The ENCODE Controversy And Professionalism In Science (this also has a Storify of some of the chatter that’s gone on via twitter, where much of this goes on these days)
This was an early one but I missed it in my travel days, by Ashutosh Jogalekar at SciAm: ENCODE, Apple Maps and function: Why definitions matter
Anshul Kundaje takes issue with some of the conclusions drawn based on the data used: https://twitter.com/anshul
Derek Lowe at In the Pipeline: ENCODE: The Nastiest Dissent I’ve Seen in Quite Some Time
Mike’s Fourth Try, Mike Lin–an author in the consortium: My thoughts on the immortality of television sets
Rebecca Boyle at PopSci: The Drama Over Project Encode, And Why Big Science And Small Science Are Different
W. Ford Doolittle Is junk DNA bunk? A critique of ENCODE
Larry Moran: Ford Doolittle’s Critique of ENCODE
Nature Editorial: Form and function
Peter, a kiwi, on The ENCODE War Continues
A collection of emails on the subject of ENCODE and our demolition paper + 2 funny pictures http://t.co/IpFu83bsBb
— Dan Graur (@DanGraur) March 20, 2013
Richard Gayle ENCODE and the Truth
(subscription req) #openaccess : The ENCODE project: Missteps overshadowing a success