I realise thus far, I may not have explained too clearly exactly what I’m doing for my Panton fellowship. With this post I shall attempt to remedy this and shed a little more light on what I’ve been doing lately.
The main thrust of my fellowship is to extract phylogenetic tree data from the literature using content mining approaches (think text mining, but not just text!) – using the literature in its entirety as my data. I have very little prior experience in this area, but luckily I have an expert mentor guiding me: Peter Murray-Rust (whom you may often see referred to as PMR). For those of us biologists who may not be familiar with his work, whilst trying not to be too sycophantic about it, PMR is simply brilliant, it’s amazing what he and his collaborators have done to extract chemical data from the chemical literature and provide it openly for everyone, in spite of fierce opposition at times from those with vested interests in this data remaining ‘closed’.
Now he’s turned his attention to the biological literature for my project and together we’re going to try and provide open tools to extract phylogenetic data from the literature. Initially I proposed trying to grab just tree topology and tip labels – a kind of bare minimum, but PMR has convinced me that we should be ambitious and all-encompassing, and thus our aims have expanded to include branch lengths, support values, the data-type the phylogeny was inferred from, and other useful metadata. And why not? We’re ingesting the totality of the paper in our process, from title page to reference list, so there’s plenty of machine-readable data to be gleaned. The question is, can we glean it off accurately enough, balancing precision and recall?
So for starters, we’ve been using test materials that we’re legally allowed to, namely Open Access CC-BY papers from BMC & PLoS to test our extraction tools, specifically focusing on a subset of all ~8500 papers containing the word-stem phylogen* from BMC. It’s a rough proxy for papers that’ll contain a tree, and it’s good enough for now – we’ll need to be able to deal with false positives along with all the positive positives, so it’s instructive to keep these in our sample.
We’ve been working on the regular structure of BMC PDFs, getting out bibliographic metadata, and the main-text for further NLP processing downstream to pick out data & method relevant words like say PAUP* , ML , mitochondrial loci etc… But the real reason we’re deliberately using PDFs rather than the XML (which we also have access to) is the figures – where all the valuable phylogenetic tree data is. If this can be re-interpreted with reference to the bibliographic metadata, the figure caption and further methodological details from the full-text of the paper, then we may be able to reconstruct some fairly rich and useful phylogenetic data.
To make it clear, in slight contrast to the Lapp et al iEvoBio presentation embedded above, we’re not trying to just extract the images, but rather to re-interpret them back into actual re-useable data, probably to be provided in NeXML (and from there on, whatever form you want). We’re pretty sure it’s an achievable goal. Programs like TreeThief, TreeRipper, and TreeSnatcher Plus have gone some way towards this already, but never before been incorporated in a content mining workflow AFAIK.
Unfortunately I wasn’t at iEvoBio 2012 (I’m short on money and on time these days) but it’s great to see from the slides the growing recognition of the SVG image file format as a brilliant tool for communicating digital science. I also put a bit about that in my Hennig XXXI talk slides too (towards the end). Programs like TNT do output SVG files, so there’s scope to make this a normal part of any publication workflow. Regrettably though, rather few publisher produced PDFs contain SVG formatted images – but if people, and editorial boards (perhaps?) can be made aware of their advantages, perhaps we can change this in future…?
The scope and scale of phylogenetic research (using ‘phylogen*’ as a proxy):
There’s a lot of phylogenetic research out there… but little of it is Open Access – which is problematic for content mining approaches – particularly if subscription-access publishers are reticent to allow access.
- with a Thomson Reuters Web of Science search, SCI-EXPANDED database (only), Topic=(phylogen*) AND Year Published=(2000-2011) this returns 101,669 results (at the time of searching YMMV)
- 91,788 of which are primary Research Articles (as opposed to Reviews, Proceedings Papers, Meeting Abstracts, Editorial Materials, Corrections, Book Reviews etc…)
- Recent MIAPA working group research I contributed to (in review) quantitatively estimates that approximately 66% of papers containing ‘phylogen*’ report a new phylogenetic analysis (new data).
- Thus conservatively assuming just one tree per paper (there are often many per paper), there are > 60,000 trees contained within just 21st century research articles.
- As with STM publishing as a whole, the number of phylogenetic research articles being published each year shows consistent year-on-year increases
- Cross-match this with publisher licencing data and you’ll find that only ~11% of phylogenetic research published in 2010 was CC-BY Open Access (and this % probably decreases as you go back before 2010)
Finally, it’s also worth acknowledging that we’re certainly not the first in this peculiar non-biomedical mining space – ‘biodiversity informaticists’ have been doing useful things with these techniques for a while now in innovative ways largely unrelated to medicine e.g. LINNAEUS from Casey Bergmann’s lab, and a recent review of other projects from Thessen et al (2012) [hat-tip to @rdmpage for bringing that later paper to the world’s attention via Twitter]. Literally all areas of academia could probably benefit from some form or another of content mining – it’s not just a biomed / biochem tool.
So, I hope that explains things a bit better. Any questions?
Some references (but not all!):
Gerner, M., Nenadic, G., and Bergman, C. 2010. LINNAEUS: A species name identification system for biomedical literature. BMC Bioinformatics 11:85+. http://dx.doi.org/10.1186/1471-2105-11-85 [CC-BY Open Access]
Thessen, A. E., Cui, H., and Mozzherin, D. 2012. Applications of natural language processing in biodiversity science. Advances in Bioinformatics 2012:1-17. http://dx.doi.org/10.1155/2012/391574 [CC-BY Open Access]
Hughes, J. 2011. TreeRipper web application: towards a fully automated optical tree recognition software. BMC Bioinformatics 12:178+. http://dx.doi.org/10.1186/1471-2105-12-178 [CC-BY Open Access]
Laubach, T., von Haeseler, A., and Lercher, M. 2012. TreeSnatcher plus: capturing phylogenetic trees from images. BMC Bioinformatics 13:110+. http://dx.doi.org/10.1186/1471-2105-13-110 [CC-BY Open Access, incidentally I was one of the reviewers for this paper. I signed my review, and made a point of it too. Nor was it a soft review either I might add]