Text mining approaches to the scientific literature have become increasingly popular as a way of helping researchers to make sense of a growing number of papers. But they've tended to focus on positive findings and skim over negative ones. In this sense they're following in the tradition of scientists themselves, unfortunately.
It's also hard to search for negative findings on PubMed, because if you type in, say, vaccines NOT associated with autism in the hopes of finding papers showing that vaccines don't cause autism, it will think you are trying to search for "vaccines" and don't want to see any papers mentioning the words "associated with autism". So you end up with 160,000 hits about vaccines with no reference to autism at all. There are ways around this but it's surprisingly tricky.
BioNOT uses text mining to mine null findings from a large database which includes everything you can find on PubMed and also a large number of full text articles (some behind paywalls).
Authors Agarwal et al of Wisconsin say that this will help to map out the "incidentalome" (a brilliant word I'd never heard before) for a given disease or trait i.e. the regions of the genome that turned out not to be associated with it. It should work for anything, though, not just genes.
However the BioNOT system isn't perfect. The authors note that it is rather over-enthusiastic in finding negative sentences.
A quick try on the system bears this out. I searched for 5 HTTLPR, the claimed "happiness gene". This revealed many papers finding no link between the gene and various things. But it also threw up false positives (how ironic), such as:
young rhesus monkeys were split into two groups... those having, or not, the short variant of the 5 -HTTLPR polymorphismThis is just telling us about the methods of a study. It's not a null finding, but it set the BioNOT alarm bells ringing, presumably because it contained the word "not".
So BioNOT is only a first step, but it's an important one.
Agarwal S, Yu H, & Kohane I (2011). BioNOT: A searchable database of biomedical negated sentences. BMC bioinformatics, 12 (1) PMID: 22032181
5 comments:
Great stuff. I came across a study earlier this year showing, at least in surgery, that negative studies were being published in second-tier journals.
(Sorry for the deleted comments, I was having trouble embedding the link.)
Finally! This is great. It's about time they start giving data on unpredictable findings (e.g. retaining the null hypothesis).
This is an amazing engine! The creators need to be showered with kudos all 'round! Thanks for the post!
This is awesome!
It's still only the negative results that make it to publication, though --- but still, awesome.
Thank you for posting about it; don't know that I would've seen this otherwise.
Hi, thanks for linking us. We are working on a new version of BioNOT which will allow users to search gene and disease names more specifically. As far as false positives (!, it was fun writing a paper with that) are concerned, it's difficult to get rid of them completely. We are trying to have a high enough signal-to-noise ratio so that the system is useful. Suggestions for improvement are welcome.
Post a Comment