Feeling negative about your PubMed searches? You're not alone. Because of the way search engines are set up, searching for correlations between, say, "gene X" and "cancer," will return lots of positive results. On the other hand, studies that find no correlation—negative or null results—are tough to find (if they're published at all). Typing the word "not" into your search only makes matters worse: "gene X NOT cancer" returns only papers in which cancer isn't mentioned at all.
Luckily, there's now an app for finding those null results: BioNOT, a new search engine developed by biomedical informatician Hong Yu and graduate student Shashank Agarwal at the University of Wisconsin, Milwaukee. The program's artificial intelligence scours PubMed for entered terms, such as "vaccines" and "autism," and returns papers with text showing the two are not related.
The goal right now, Yu says, is to help with gene annotation efforts and allow search engines to better pinpoint whether or not current research links a condition to a gene. Complex cancers, for instance, may involve hundreds of genetic variants; looking at all the papers on all of them together, she says, can mask actual findings and weaken the statistical strength of the correlation. And, perhaps more importantly, different labs that study a gene get different results. Searching PubMed and BioNOT in parallel could give researchers two lists of papers that find positive and negative results, allowing them to easily compare the strength of the data in each list.
In a paper the team published in BMC Bioinformatics last week, the researchers used BioNOT to search for genes thought to be involved in three diseases: autism, Alzheimer's, and Parkinson's. The tool dug up negative results that otherwise may have been missed, the authors say. Right now, BioNOT is limited to full texts from mega-publisher Elsevier (which Yu says is interested in investing in further development of BioNOT), open-access publications, and PubMed abstracts. This gives the tool about 32 million negated sentences to work with, and Yu hopes that more journals will allow BioNOT to access their full text it in the future.
Literature curators who annotate genomes or "text mine" scientific literature are interested in the application. "There are a lot of text mining applications out there, but this one is quite unique," says Emily Dimmer, director of the UniProt GO protein annotation program at the European Bioinformatics Institute (EBI) in the United Kingdom. EBI, she says, has just started to use BioNOT to better focus their data collection and protein annotation efforts.
The program still has some blips: searching for correlations between cancer and cell phones, for instance, pulls up a number of papers that contain negated sentences but have nothing to do with the correlation between the two. But Yu says that they plan to enhance this aspect of the program. And the researchers are encouraged by interest from both publishers and bioinformaticians such as EBI. "It's not just enjoyable, but also rewarding work," she says, since it can enhance biomedical discovery. "We're at a time where AI [artificial intelligence] is very powerful."