Community Annotation - September 2010 Satellite
Community Annotation Satellite Meeting
September 2010 GMOD Meeting
15 September 2010
This satellite meeting was held Wednesday, 15 September, in Cambridge, UK as part of GMOD Europe 2010 and the September 2010 GMOD Meeting. This session dealt with Community Annotation. The other satellite was on Post Reference Genome Tools and had many of the same participants. The topic of community annotation was proposed by Kim Rutherford.
- Ellen Adlem, Cambridge University Cambridge Institue of Medical Research, T1DBase
- Gerd Anders, Max-Delbrueck-Centrum Berlin (MDC)
- Jerven Bolleman, UniProt Swiss-Prot
- Scott Cain, GMOD Project Coordinator, OICR
- Dave Clements, GMOD Help Desk, NESCent
- Oskana Riba Grognuz, Swiss Institute of Bioinformatics (SIB) Department of Ecology and Evolution, University of Lausanne
- Chris Hemmerich, CGB, Indiana University
- Joan Pontius, National Cancer Institute, SAIC
- Kim Rutherford, Cambridge Systems Biology Centre, PomBase
Val Wood at PomBase sent out a GO curation request in a Microsoft Word document to 80-90 authors of recent S. pombe papers. Val is well known in the community and asked the authors to provide GO annotation for their papers. PomBase got an 80% response rate. The annotations were reviewed by PomBase personnel and most annotations were either spot on, or very close in the GO hierarchy. Val's view is if PomBase can get the community to do 90% of the annotation work, then PomBase can do 10 times as much annotation.
UniProt added a "Contribute" section at the top of their protein pages. It includes 2 links, "Send feedback" and "Read comments or add your own." In the year it has been up, UniProt has received 9 comments. That's less than one comment a month for a resource that averages almost 4 million page views per month. The "Send feedback" link gets used more often, but not often enough to be a serious time commitment for UniProt staff.
Curators at UniProt also contact authors of a paper when the paper is added to UniProt. This often results in the author's coming back to UniProt more often.
ZFIN, the zebrafish model organism database (MOD), was launched in the mid 1990s with the stated goal of enabling community annotation. The development team had a user interface background and went to great lengths to support this. ZFIN's annotation concept was for individuals to own certain gene and fish pages. This ownership gave them editing privileges for those pages. This effort was largely unsuccessful and today the vast majority of ZFIN users only have update access to their lab and personal records at ZFIN.
ZFIN has a 'Your Input Welcome' button on every data page. An average of 6 comments are received per month. These comments include questions, corrections, additional data, and requests to back curate data from older publications. ZFIN has had success with teams of community experts assisting in the development of the zebrafish anatomical ontology. In recent years, ZFIN, like many MODs, has added a community wiki. They are specifically encouraging community members to add protocols and antibody information. However, despite being more convenient and more familiar to users than ZFIN's original efforts, there has been viewing traffic, but no user updates so far.
ZFIN encourages labs to submit gene expression and phenotype data using the Phenote software tool. Phenote allows researchers to annotate their images using terms from the ontologies used by ZFIN curators. There have been inquiries but there have not been any submissions with this tool to date.
MGI and EMAGE
Both MGI, the mouse MOD, and EMAGE, the Edinburgh Mouse Atlas of Gene Expression launched tools aimed at helping mouse researchers keep track of their gene expression data. These two projects took very different approaches, but ended up with similar, disappointing results. MGI created the Gene Expression Notebook (GEN), based on Excel, that bench scientists could use to track gene expression experiments and record annotation. This tool addressed a common need (tracking gene expression data and metadata) with a technology (Microsoft Excel) that almost every biologist is already familiar with. Researchers could use GEN for their own work, and then optionally submit their annotations back to MGI if they wanted to.
EMAGE launched a desktop Java application for doing image based gene expression annotation. The goal was similar to MGI's but the tool allowed researchers to do annotation at much finer detail. However, it also required them to learn the ins and outs of an entirely new interface.
Several resources, including ParameciumDB and the Pea Aphid database, have offered training in genome annotation using Apollo. Apollo is a sophisticated tool that allows users to add evidence and refine gene predictions and other annotations. However, some efforts have hit bumps along the way with users who don't use the tools often enough to remember the user interface between annotation sessions.
The CACAO project at Texas A&M teaches GO annotation to undergraduates and then adds the resulting annotations to EcoliWiki. This program has produced a relatively large volume of quality annotation by using several techniques to specifically encourage this. Students are divided into teams that compete with each other for points based on the quantity and quality of annotation. Quantity is promoted through offering points for number of annotations and quality by offering points for successfully challenging other teams' faulty annotations. Winning teams are rewarded with pizza and soft drinks. This has been much more successful than expected, and is a win for both the students and the resources the annotations go into. This approach has been a stunning success with an order of magnitude more annotations coming in than organizers expected.
ParameciumDB also harnesses the power of undergraduates to do genomic annotation. They have partnered with two undergraduate institutions that teach term-long classes on annotation. Students annotate the paramecium genome, and then submit their work to ParameciumDB at the end of the term. This work is then reviewed and loaded into ParameciumDB.
The Science Education Alliance (SEA) at HHMI takes a similar approach on a broader scale, starting with gathering and sequencing samples through submission to GenBank.
EcoliWiki is a recent effort that aggressively encourages community annotation. EcoliWiki is built on the widely used MediaWiki package (MediaWiki powers Wikipedia.org) and exposes all it's annotation to being updated and expanded by anyone with a login.
CACAO and EcoliWiki make extensive use of wiki technology to make the technology aspects of entering annotation easy for researchers and students. EcoliWiki's update interface makes extensive use of the the locally developed, but publicly available, TableEdit MediaWiki extension. This presents editors with a GUI interface to tabular data and protects them from MediaWiki markup.
Wikis offer significant hope for making it easier for biologists to fix or add something than it is to be irritated by an error or the absence of it.
Access and Logins
What model should be used for enabling and managing community annotation. Who do you give access to to, what types of access do they have, and what review mechanisms are in place?
- Wikipedia model - Last edit wins.
- Anyone can edit, with or without a login. Logins are available to anyone who requests it. Spam is automatically removed, but differences of opinion are resolved by last edit wins in most cases, and by increasing levels of control when that doesn't work.
- GMOD.org - All edits welcome, but you'll be watched.
- Must have a login to edit, but anyone can create a login. Edits are loosely watched by GMOD staff.
- EcoliWiki - Vampire!
- Logins given to trusted community members. Anyone with an existing login can create a login for someone else. Editors have full update access.
- SGN - Locus boss!
- SGN locus pages list the set of community members who can update information about that locus. Every locus page includes "Request editor privileges" link. This model allows specific people to control and "own" particular loci data. Locus owners tend to be experts on that locus, and giving control to them (and listing their names) benefits both SGN and the locus editor.
- PomBase (so far) - Reviewed before posting.
- Submissions are reviewed by staff at the online resource before being posted to the public resource.
- ParameciumDB - Vetting, followed by complete trust
- ParameciumDB only give access to people who know how to annotate. These annotations are not reviewed before incorporation. They are incorporated automatically once a month if they are tagged as finished by the community curator.
We also discussed problems with requiring researchers to create yet another online login.
EcoliWiki automatically creates logins for authors of newly enterred papers. The login is then sent to authors when they are first contacted. Another option could be to send them a custom URL that allows them to update the resource without having to login. If they access that URL, they are who they say they are.
(Q: Is anyone out there investigating unified logins in life sciences? This could be done as a front end to OpenID?)
Attribution and Credit
Want all statements in your resource to be attributed. This means you need field level (not just page level) attribution. Who said this, where it was said. With community annotation, you want to record that information, plus who put that information in the database. ZFIN tracks all curator changes at the column/attribute level. Users can view this history for any web page.
The lack of "professional" credit for doing community annotation has been an enormous issue in the past. People are becoming aware that online resources help other researchers find their work. There has been talk of shifting funding and tenure models to consider contributions to online resources. (If you know of any concrete examples of where this has been done, please post them here.)
One participant pointed out that students could be encouraged to do community annotation. This would enable them to cite resources online that they had helped create. This sort of meta-publication is not original research but it would demonstrate that a student is able to read and understand academic papers.
Level of Detail
Resources need to be careful not to overwhelm their community members. Dedicated curators can learn sophisticated tools and doing highly detailed annotation. Community members who may want to submit or update information only 3 or 4 times year, are not inclined to learn (or remember) complex tools or ontologies with thousands of terms. Community annotation tools, unlike dedicated curator tools must be easy and intuitive to use.
One approach is to complex, large ontologies is to use reduced vocabularies such as GO slim. Have tools default to GOSlim, and then only once they have selected a GO-Slim term, prompt them for a narrower term from the full ontology, if they want to provide one.
PomBase is going to create a lightweight GO annotation tool that autocompletes as much as possible. This will create a GAF, but won't require users to manually populate every column. The same applies to anatomy - how much does your average researcher care about 600 terms in a mouse kidney? Unless they are a kidney researcher that level of detail may overwhelm users.
Despite the complexity issues that come with ontologies, everyone present agree that abandoning ontologies in favor or free text annotation would be a huge mistake.
Data Quality and Consistency
Problems of consistent annotation happen even when resources have a dedicated staff of in-house curators. Early in the GUDMAP project, two highly qualified curators did independent parallel annotations of the same in situ hybridisation image dataset. Initially, well over half of the annotation agreed completely. The two curators then met and identified the sources of disagreement. Several areas were clarified and the datasets reannotated, at which point two reached 96% agreement. However, despite being experts in this field, the two still disagreed on what 4% of the image were showing.
This problem becomes more pronounced with community annotation. Community members won't create annotation full time or even with a significant fraction of their time. Most resources won't be able to teach community members best practices. Resources will need to decide on a model for what annotation they will encourage and accept.
However, community annotation can also help increase data quality. One resource represented in the discussion believes that on average, 3% of their data is incorrect. Databases can use their communities to help identify incorrect data by making it trivial to report problems and easy to fixt hem. You can also use your community to resolve conflicting annotations.
Can we get journals to require supplemental annotation data with publications? This very topic was discussed at a GMOD meeting 6 or 7 years ago. Everyone agreed that it was a good idea then too.
Communities Without a MOD
What can be done for communities that don't have a central resource for the gathering and itegration of data? GONUTS is a resource for GO annotation for any organism. Other data-type specific resources are available for such communities. However, does anything exist for these communities that is unified?
Why were some of these efforts more successful than others?
The PomBase example above was wildly successful compared to most community annotation efforts. It has some beneficial features that can be controlled by MODs, and some that cannot: PomBase contacted the author's of recent papers only. Authors are much more likely to care about research that has been published this year, then they are about previous work. The S. Pombe community is also a small and relatively well connected community. Community annotation efforts tend to work better in smaller communities where researchers are likely to know and work with the people who will benefit from the annotation. Adding annotation to one's work is also likely to significantly increase the visibility of that work in smaller communities. Increased visibility is likely to lead to increased citations. In large communities, such as mouse or yeast, adding annotation is likely to only lead to an incremental increase in visibility and citations.
ParameciumDB has also had some success in encouraging community annotation. The paramecium community, like the pombe community, is also small and collaborative. UniProt, a pan-biology resource if ever there was one, had 46 million page views lead to only 9 comments being submitted. ZFIN likely suffered from being both a large community and from being an early adopter.
Neither of the mouse gene expression tools led to significant amounts of data being submitted back to the organizations that created them. The suspicion is that, for MGI's GEN, every lab does things slightly differently, and has already developed processes for doing this. The payoff for embracing a tool that enables data sharing is not perceived as being high enough to change processes. For EMAGE, it may be that the perceived benefits of doing detailed, high-quality, image-based annotation were perceived to be not worth the effort to learn a non-trivial tool. This problem will occur anytime a tool does a complex task and thus has a complex interface. This problem will, however, lessen over time, as tools move from custom desktop applications to web based one, using the same user interface conventions as other sophisticated interfaces. Apollo, for example, is currently being migrated from a stand-alone Java application to web application that shares a common platform with JBrowse. This common platform will ease adaptation of Apollo by any user that is already familiar with JBrowse.
ParameciumDB and EcoliWiki's CACAO project, while not being community annotation efforts per se, are successful at generating significant amounts of curation by people who are not paid curators. These are both win-win situations as the students learn about curation and the resources get much needed annotation.
CACAO is also be taking advantage of a culture shift. The students are already used to creating online content that others will see. This was exceptional when ZFIN launched in the mid 1990s. Now it is commonplace.
Some possible lessons from the above experience and analysis.
- Community size matters
- Community annotation efforts tend to work better in smaller communities, such as PomBase and ParameciumDB. In larger communities such as mouse or zebrafish development, community members feel that there are organizations with full time curators such as MGI and ZFIN that exist to do this annotation for them. In both large and small communities there is pressure to publish, but in smaller communities the perceived benefits of doing community annotation are greater. Large pan-biology resources such as UniProt have the greatest challenge in encouraging community contributions.
- Contact researchers and authors when their research first comes out.
- They are interested and are keen to help increase the visibility of their work.
- Contact authors when papers and research are new, or immediately after or just before adding their work to a website. Authors are much more likely to respond when the research is fresh and added annotation has the potential to significantly raise the visibility of their work. This contact also increases the likelihood that the authors will use the resource.
- Complexity and interface familiarity matters
- There is often a tradeoff between rich functionality and ease of use. This will likely get better over time as web interfaces become both more sophisticated and more standardised.