Show me the data!
Header

To prove my point about the way that supplementary data files bury useful data, making it utterly indiscoverable to most, I decided to do a little experiment (in relation to text mining for museum specimen identifiers, but also perhaps with some relevance to the NHM Conservation Hackathon):

I collected the links for all Biology Letters supplementary data files. I then filtered out the non-textual media such as audio, video and image files, then downloaded the remaining content.

A breakdown of file extensions encountered in this downloaded subset:

763 .doc files
543 .pdf files
109 .docx files
75 .xls files
53 .xlsx files
25 .csv files
19 .txt files
14 .zip files
2 .rtf files
2 .nex files
1 .xml file
1 “.xltx” file

I then converted some of these unfriendly formats into simpler, more easily searchable plain text formats:

 

Now everything is properly searchable and indexable!

In a matter of seconds I can find NHM specimen identifiers that might not otherwise be mentioned in the full text of the paper, without actually wasting any time manually reading any papers. Note, not all the ‘hits’ are true positives but most are, and those that aren’t e.g. “NHMQEVLEGYKKKYE” are easy to distinguish as NOT valid NHM specimen identifiers:

 

Perhaps this approach might be useful to the PREDICTS / LPI teams, looking for species occurrence data sets?

I don’t know why figshare doesn’t do deep indexing by default – it’d be really useful to search the morass of published supplementary data that out there!

Progress on specimen mining

June 14th, 2015 | Posted by rmounce in Content Mining - (0 Comments)

I’ve been on holiday to Japan recently, so work came to a halt on this for a while but I think I’ve largely ‘done’ PLOS ONE full text now (excluding supplementary materials).

My results are on github: https://github.com/rossmounce/NHM-specimens/tree/master/results – one prettier file without the exact provenance or in-sentence context of each putative specimen entity, and one more extensive file with provenance & context included which unfortunately github can’t render/preview.

 

Some summary stats:

I found 427 unique BMNH/NHMUK specimen mentions from a total of just 69 unique PLOS ONE papers. The latter strongly suggests to me that there are a lot of ‘hidden’ specimen identifiers hiding out in difficult-to-search supplementary materials files.

I found 497 specimen mentions if you include instances where the same BMNH/NHMUK specimen is mentioned in different PLOS ONE papers.

Finding putative specimen entities in PLOS ONE full text is relatively automatic and easy. The time-consuming manual part is accurately matching them up with official NHM collection specimens data.

I could only confidently link-up 314 of the 497 detected mentions, to their corresponding unique IDs / URLs in the NHM Open Data Portal Collection Specimens dataset. Approximately one third can’t be confidently be matched-up to a unique specimen in the online specimen collection dataset — I suspect this is mainly down to absence/incompleteness in the online collections data, although a small few are likely typo’s in PLOS ONE papers.

In my last post I was confident that the BM Archaeopteryx specimen would be the most frequently mentioned specimen but with more extensive data collection and analysis that appears now not to be true! NHMUK R3592 (a specimen of Erythrosuchus africanus) is mentioned in 5 different PLOS ONE papers. Pleasingly, Google Scholar also finds only five PLOS ONE papers mentioning this specimen – independent confirmation of my methodology.

One of the BM specimens of Erythrosuchus is more referred to in PLOS ONE than the BM Archaeopterx specimen

Now I have these two ‘atomic’ identifiers linked-up (NHM specimen collections occurrence ID + the Digital Object Identifier of the research article in which it appears), I can if desired, find out a whole wealth of information about these specimens and the papers they are mentioned in.

My next steps will be to extend this search to all of the PubMedCentral OA subset, not just PLOS ONE.

 

In this post I’ll go through an illustrated example of what I plan to do with my text mining project: linking-up biological specimens from the Natural History Museum, London (sometimes known as BMNH or NHMUK) to the published research literature with persistent identifiers.

I’ve run some simple grep searches of the PMC open access subset already, and PLOS ONE make up a significant portion of the ‘hits’, unsurprisingly.

Below is a visual representation of the BMNH specimen ‘hits’ I found in the full text of one PLOS ONE paper:

Grohé C, Morlo M, Chaimanee Y, Blondel C, Coster P, et al. (2012) New Apterodontinae (Hyaenodontida) from the Eocene Locality of Dur At-Talah (Libya): Systematic, Paleoecological and Phylogenetical Implications. PLoS ONE 7(11): e49054. doi: 10.1371/journal.pone.0049054

onepaper

I used the open source software Gephi, and the Yifan Hu layout to create the above graphical representation. The node marked in blue is the paper. Nodes marked in red are catalogue numbers I couldn’t find in the NHM Open Data Portal specimen collections dataset: 10 out of 34 not found.

Source data table below showing how uninformative the NHM persistent IDs are. I would have plotted them on the graph instead of the catalogue strings as that would be technically more correct (they are the unique IDs), but it would look horrible.

 

I’ve been failing to find a lot of well known entities in the online specimen collections dataset which makes me rather concerned about its completeness. High profile specimens such as Lesothosaurus “BMNH RUB 17” (as mentioned in this PLOS ONE paper, Table 1) can’t be found online via the portal under that catalogue number. I can however find RUB 16, RUB 52 and RUB 54 but these are probably different specimens. RUB 17 is mentioned in a great many papers by many different authors so it seems unlikely that they have all independently given the specimen an incorrect catalogue number – the problem is more likely to be in the completeness of the online dataset.

Another ‘missing’ example is “BMNH R4947” a specimen of Euoplocephalus tutus as referred to in Table 4 of this PLOS ONE paper by Arbour and Currie. There are two other records for that taxon, but not under R4947.

To end on a happier note, I can definitely answer one question conclusively:
What is the most ‘popular’ NHM specimen in PLOS ONE full text?

…it’s “BMNH 37001”, Archaeopteryx lithographica which is referred to in full text by four different papers (see below for details).

I have feeling many more NHM specimens are hiding out in separate supplementary materials files. Mining these will be hard unless figshare gets their act together and creates a full-text API for searching their collection – I believe it’s a metadata only API at the moment.

37001 in PLOS ONE papers

 

I’ve purposefully made very simple graphs so far. Once I get more data, I can start linking it up to create beautiful and complex graphs like the one below (of the taxa shared between 3000 microbial phylogenetic studies in IJSEM, unpublished), which I’m still trying to get my head around. The linked open data work continues…

Bacteria subutilis commonly used

 

Now I’m at the Natural History Museum, London I’ve started a new and ambitious text-mining project: to find, extract, publish, and link-up all mentions of NHM, London specimens published in the recent research literature (born digital, published post-2000).

Rod Page is already blazing a trail in this area with older BHL literature. See: Linking specimen codes to GBIF & Design Notes on Modelling Links for recent, relevant posts. But there’s still lots to be done I think, so here’s my modest effort.

 

Why?

It’s important to demonstrate the value of biological specimen collections. A lot of money is spent cataloguing, curating and keeping safe these specimens. It would be extremely useful to show that these specimens are being used, at scale, in real, recent research — it’s not just irrelevant stamp collecting.

Sometimes the NHM, London specimen catalogue has incorrect, incomplete or outdated data about it’s own specimens – there is better, newer data about them in the published literature that needs to be fed back to the museum.

An example: specimen “BMNH 2013.2.13.3” is listed in the online catalogue on the NHM open data portal as Petrochromis nov. sp. By searching the literature for BMNH specimens, I happened to find where the new species of this specimen was described: http://dx.doi.org/10.1007/s10228-014-0396-9 as Petrochromis horii Takahashi & Koblmüller, 2014. It’s also worth noting this specimen has associated nucleotide sequence data on GenBank here: http://www.ncbi.nlm.nih.gov/nuccore/AB850677.1 .

Having talked a lot about the 5 stars of open data in the context of research data recently, I wonder… wouldn’t it be really useful to make 4 or 5 star linked open data around biological specimens? From Rod Page, I gather this is part of the grand goal of creating a biodiversity knowledge graph.

For this project, I will be focussing on linking BMNH (NHM, London) specimen identifiers with publication identifiers (e.g. DOIs) and GenBank accession numbers.

 

What questions to ask?

Where have NHM, London specimens been used/published? What are the most used NHM, London specimens in research? How does NHM, London specimen usage compare to other major museums such as the AMNH (New York) or MNHN (Paris).

Materials for Mining

1.) The PubMedCentral Open Access subset – a million papers, but mainly biomedical research.
2.) Open Access & free access journals that not included in PMC
3.) figshare – particularly useful if nothing else, as a means of mining PLOS ONE supplementary materials (I read recently that essentially 90% of figshare is actually PLOS ONE supp. material! See Table 2)
4.) select subscription access journals – annoyingly hard to get access to in bulk, but important to include as sadly much natural history research is still published behind paywalls.

 

(very) Preliminary Results

The PMC OA subset is fantastic & really facilitates this kind of research – I wish ALL of the biodiversity literature was aggregated like (some) of the open access biomedical literature is. You can literally just download a million papers, click, and go do your research. It facilitates rigorous research by allowing full machine access to full texts.

Simple grep searches for ‘NHMUK’ & ‘BMNH [A-Z0-9][0-9]’, two of the commonest citation forms by which specimens may be cited reveal many thousands of possible specimen mentions in the PMC OA subset I must now look through to clean-up & link-up. In terms of journals, these ‘hits’ in the PMC OA subset come from (in no particular order): PLOS ONE, Parasites & Vectors, PeerJ, ZooKeys, Toxins, Zoo J Linn Soc, Parasite, Frontiers in Zoology, Ecology & Evolution, BMC Research Notes, Biology Letters, BMC Evolutionary Biology, Aquatic Biosystems, BMC Biology, Molecular Ecology, Journal of Insect Science, Nucleic Acids Research and more…!

specimen “BMNH 86.10.4.2” is a great example to lookup / link-up on the NHM Open Data Portal: http://data.nhm.ac.uk/specimen/8559613e-f2a3-447c-aa1a-d476600d3293 the catalogue record has 7 associated images openly available under CC BY, so I can liven up this post by including an image of the specimen (below)! I found this specimen used in a PLOS ONE paper: Walmsley et al. (2013) Why the Long Face? The Mechanics of Mandibular Symphysis Proportions in Crocodiles. doi: 10.1371/journal.pone.0053873 (in the text caption for figure 1 to be precise).

© The Trustees of the Natural History Museum, London. Licensed for reuse under CC BY 4.0. Source.

 

 

Questions Arising

How to find and extract mentions of NHM, London specimens in papers published in Science, Nature & PNAS ? There are sure to be many! I’m assuming the last 15 years worth of research published in these journals will be difficult to scrape – they would be quite likely to block my IP address if I tried to. Furthermore, all the actual science is typically buried in supplementary file PDFs in these journals not in the ‘main’ short article. Will Science, Nature & PNAS  let me download all their supp material from the last 15 years? Is this facilitated at all? How do people actually do rigorous research when the contents of supplementary data files published in these journals are so undiscoverable & inaccessible to search?

 

It’s clear to me there are many separate divisions when it comes to discoverability of research. There’s the divide between open access (highly discoverable & searchable) and subscription access (less discoverable, less searchable, depending upon publisher-restrictions). There’s also the divide between the ‘paper’ (more searchable) and ‘supplementary materials’ (less easily searchable). Finally, there’s also the divide between textual and non-textual media: a huge amount of knowledge in the scientific literature is trapped in non-textual forms such as figure images which simply aren’t instantly searchable by textual methods (figure captions DO NOT contain all of the information of the figure image! Also, OCR is time consuming and error-prone especially on the heterogeneity of fonts and orientation of words in most figures). For example, looking across thousands of papers with phylogenetic analyses published in the journal IJSEM, 95% of the taxa / GenBank accessions used in them are only mentioned in the figure image, nowhere else in the paper or supplementary materials as text! This needs to change.

 

As should be obvious by now; this is a very preliminary post, just to let people know what I’m doing and what I’m thinking. In my next post I’ll detail some of the subscription access journals I’ve been text mining for specimens, and the barriers I’ve encountered when trying to do so.

 

Bonus question: How should I publish this annotation data?

Easiest would be to release all annotations as a .csv on the NHM open data portal with 3 columns where each column mimics ‘subject’  ‘predicate’ ‘object’ notation: Specimen, “is mentioned in”, Article DOI.

But if I wanted to publish something a little better & a little more formal, what kind of RDF vocabulary can I use to describe “occurs in” or “is mentioned in”. What would be the most useful format to publish this data in so that it can be re-used and extended to become part of the biodiversity knowledge graph and have lasting value?