This presentation was given at the CLIR/DLF Postdoctoral Fellowship Summer Seminar at Bryn Mawr college in Pennsylvania on July 29th 2014. The intention was to communicate what we are doing in the fields of text and data mining in the domain of chemistry and specifically around mining the RSC archive publication and chemistry dissertations and theses. How would these experiences map over to the humanities?
Thyroid Physiology_Dr.E. Muralinath_ Associate Professor
Data Mining Dissertations and Adventures and Experiences in the World of Chemistry
1. Data mining dissertations
Adventures and Experiences in the World of Chemistry
Antony Williams
CLIR/DLF Postdoctoral Fellowship Summer Seminar,
July 2014
3. • Who’s got an ORCID?
• Who has heard of/involved with AltMetrics?
• Who has edited a Wikipedia page?
• Who has direct experience of text mining?
• All slides already on Slideshare here:
• www.slideshare.net/AntonyWilliams
Before we start….
4. • Context – why do we want to mine data?
• Our experiences in extracting theses:
– Text and data mining
– Chemistry as an example
– Before you start
– Resources and tools
Contents
5. • Let’s map together all historical chemistry
data and build systems to integrate
• Heck, let’s integrate chemistry and biology
data and add in disease data too
• Let’s model the data and see if we can
extract new relationships – quantitative and
qualitative
• Let’s make it all available on the web
Taking on a big challenge…
6.
7. • We’re going to map the world
• We’re going to take photos of as many
places as we can and link them together
• We’ll let people annotate and curate the map
• Then let’s make it available free on the web
• We’ll make it available for decision making
• Put it on Mobile Devices, Give it Away
What about this….
17. • So the world can be mapped…
• We can enter a 3D world within the map
• We can add annotations
• We can use the data, reference it, we can
extract it, we can make decisions with it
• And we can do it on our lap, in our hands
• Let’s do this for chemistry…
Whoa…
18. • Once upon a time we built a database….
In a basement not far away…
23. • This is not new, you known the story…
• So much data of value contained within a
publication and delivered in a PDF form
• “PDF files, and especially unclear licensing,
don’t allow me at the data so I can rework,
reuse, repurpose, text mine etc.”
• “I specialize in XXXX. I want a database of
YYYY extracted from publications and made
available, for free, with capabilities I need,
and the publishers should just do it”
Data in a Scientific Publication
24. It is so difficult to navigate…
What’s the
structure?
Are they in
our file?
What’s
similar?
What’s the
target?Pharmacology
data?
Known
Pathways?
Working On
Now?Connections
to disease?
Expressed in
right cell type?
Competitors?
IP?
25. • Manage “all” of the chemistry data associated
with chemical substances
• Data to be downloadable, reusable, interactive
• Build a platform that enables the scientist
• Data storage, validation, standardization and
curation
• Collaborative data sharing
• Provide data platform that can enable and
enhance publishing of scientific papers
We set a vision…
26. • Every compound from every article at RSC is
extracted, in a database, and linked
• Chemical properties are extracted, databased
and used for predictive models
• Data tables are downloadable, interactive and
not just “dumb-PDFs”
• …and what can we extract from chemistry
theses too?
XXX Years from Now at RSC
27. • We are seen as one of the repositories for
published AND unpublished research data
• An intuitive platform for research data
management in the cloud
• Individual, collaborative and public data
management of diverse data in the cloud
• …and where all data referenced in a
thesis is available at a button click
XXX Years from Now at RSC
28. • But how does it map onto your domain??
So this is chemistry…
30. • You have a mountain of stuff which
contains valuable nuggets
• You (more or less) know what you’re
looking for
• You know what you’re going to do with it
once you have it
Mining as an allegory - intent
31. • You get lots of stuff out
• It requires sifting and grading
• It’s a triumph if you manage to extract
80-90% of what is there
• You will go back to the heap and redo it
Mining as an allegory - result
32. • That which is easy to get out - is well
known and unlikely to be novel
• The novel and interesting stuff is likely to
be rare and not easily defined
Mining as an allegory - effort
33. • Do the initial investigations by hand
• Send in the machines later
• Still needs some humans tweaking
Mining as allegory - automation
35. • From Utopia Documents team
• Good at extracting structure from typeset pdfs
• http://pdfx.cs.man.ac.uk/
PDFX
36. OCR recognition
• Underlining doesn’t help OCR
• In this case it was the only signpost to the
department, supervisor and funding details
37. • Hardcopy
• Scanned and OCR’d PDF
• PDF derived from Word
• Word or LaTeX
• …and for OCR not all are borne equal
• …and of course history and language is
a major influence. “Oil of vitriol”
Building blocks to mine…
38. • Ontologies, taxonomies, dictionaries
• But these are very domain focussed…
• As an example, Open PHACTS spend a
lot of effort mapping biology to chemistry
to disease over many data sources
More building blocks
39.
40. • Provide a controlled vocabulary – what
your data describes, where it came from
• Provide a shared vocabulary for
integrating with other people’s data
What can ontologies do for me?
41. Questions to ask:
(1)Has someone already produced an ontology
covering your area? (Places to look:
Bioportal, OBO Foundry.)
(2)Do they take requests?
(3)Are they responsive?
(4)Is the ontology kept up to date?
Early days for ontologies and any ontology will
almost certainly be a long way from complete!
Best practices: experiences
from biomedical ontologies
42. • Best that these don’t change
• Best that everyone calls them the same things
• Best that they are unambiguous
• Meanwhile, back in the real world
What things are you looking for?
43. • Place names – somewhat ambiguous
• Species names – can change with time
• Diseases – every pharmaceutical company
has a different list
• People – can be very ambiguous: Authors
and researchers are hard to map…except
for Google it seems!
How easy?
49. • All publications easily connected but also
– Important in early scientific career –
consider every data point contribution, every
“research object”
– Every article
– Every presentation
– Thesis and dissertation
– Provenance….and feeding AltMetrics
So the benefits of ORCIDs?
57. Tinman - mutant fly embryos lack a heart.
Van Gogh - hair-like bristles on wings have a swirling pattern.
INDY - acronym for I'm Not Dead Yet, they live twice as long as normal; from the scene in
the movie "Monty Python and the Holy Grail"
Ken and Barbie - males and females lack external genitalia.
Tribbles - some cells divide uncontrollably
Cheap date - flies are extra-sensitive to alcohol.
Cleopatra - flies die when Cleopatra gene interacts with another gene, Asp.
Kojak - no hairs on wings.
Maggie - fly development is arrested; named after Maggie Simpson, who's development
also seems to be arrested.
Oh my..Fruitfly gene names
• http://stlists.blogspot.co.uk/2005/05/fruitfly-gene-names.html
58.
59. • those that belong to the Emperor,
• embalmed ones,
• those that are trained,
• suckling pigs,
• mermaids,
• fabulous ones,
• stray dogs,
• those included in the present classification,
• those that tremble as if they were mad,
• innumerable ones,
• those drawn with a very fine camelhair brush,
• others,
• those that have just broken a flower vase,
• those that from a long way off look like flies.
Allegedly from “Celestial Emporium of Benevolent Knowledge”
The Analytical Language of John Wilkins, Jorge Luis Borges
Animal classification
60. • Are you just identifying entities?
• Are you looking for sentiment?
• In chemistry names will lead you to a
recipe for synthesis, and analytical data
about that compound
Classification after “things”
61. • Used to aid discovery - directly
• Used to aid discovery - indirectly
• Extract data in electronic form for reuse
• Needs to be use case driven – why, then
what/how comes later
End result
62. • Automation can give good results
• Especially looked at in bulk
• Less easy to judge at the article level
• People accept discovery is fuzzy
• Not so with data points
• (but maybe can screen out)
Quality
63. • Chemical names are both difficult and rewarding.
• Difficult in the sense that they can break
standard software.
• Rewarding in the sense that you can extract
useful information about the molecule they’re
referring to without a dictionary.
• Some examples…
Chemistry-specific challenges
and opportunities
70. A series of mono and di-N-2,3-epoxypropyl N-
phenylhydrazones have been prepared on a large scale
by reaction of the corresponding N-phenylhydrazones of
9-ethyl-3-carbazolecarbaldehyde, 9-ethyl-3,6-
carbazoledicarbaldehyde, 4-dimethyl-amino-, 4-
diethylamino-, 4-benzylethylamino-, 4-(diphenylamino)-,
4-(4,4-4′-dimethyl-diphenylamino)-, 4-(4-
formyldiphenylamino)- and 4-(4-formyl-4′-
methyldiphenyl-amino)benzaldehyde with
epichlorohydrin in the presence of KOH and anhydrous
Na(2)SO(4).
From Molecules, via the BioNLP list
Annotate this...
71. How many explicit compounds?
• How many numbered compounds
actually are named in a given
paper?
• iloprost (1)
• tributyl-1-hexynylstannane (2)
• the desired 2-heptyne (3)
• methyl–Pd(II) iodide 4 or 4′
• alkynylstannane 5
• the hypervalent stannate 6
• (alkynyl)(methyl)Pd(II) complex 7
• the desired methylalkyne 8
• compounds 9–14
• the stannyl precursors 15 and 16
• methylated compounds 17 and 18
• stannyl precursor 19
• iloprost methyl ester 20
• “iloprost methyl ester” is the real
name, but you need to know that
iloprost is a monocarboxylic acid!
72.
73. Names from structures
• Systematic names can be generated FROM
chemical structures algorithmically
74. General-purpose parsers do
NOT get chemical names
Visualization by bpodgursky.com using d3.js; parsing by
Stanford’s CoreNLP.
76. • OPSIN (chemical name to structure)
http://opsin.ch.cam.ac.uk/
Tools to try
77. Not all names are systematic..
Antony Williams vs Identifiers
Passport ID
Dad, Tony, others
SSN
Green Card
License
5 email addresses
ChemSpiderman (blog,
Twitter account,
Facebook, Friendfeed)
OpenID
….
83. • All of the tasks below are possible to varying extents.
Pioneered on journal abstracts and journal full text.
– Named entity recognition: what is this about? Where are
the places mentioned? Who are the people?
– Clustering and classification: which other dissertations
are like this one? What genres of dissertations are there?
– Event extraction: what processes (chemical reactions,
gene expression) occur? What are the participants?
– Citation analysis: who do dissertations cite?
– What sentiments towards the citations do authors express?
Dissertation analysis
85. • Probably less structured than papers
• Not much work has been done here before
Dissertation specifics
86. • For example:
• Stylometrics (to find out who wrote this)
• Language identification
• What else?....
• In addition to above, there are different tasks
we can perform on scientific publications and
dissertations
Digital Humanities textual
analysis tasks
87. • We would LOVE to bring data out of our
archive
• What could we do?
• Find chemical names and generate structures
• Find chemical images and generate structures
• Find reactions – and make a database!
• Find data (MP, BP, LogP) and host. Build
models!
• Find figures and database them
• Find spectra (and link to structures)
• Validate the data algorithmically
“Data enable” publications?
89. Text Mining
The N-(β-hydroxyethyl)-N-methyl-N'-(2-trifluoromethyl-1,3,4-
thiadiazol-5-yl)urea prepared in Example 6 , thionyl chloride
( 5 ml ) and benzene ( 50 ml ) were charged into a glass
reaction vessel equipped with a mechanical stirrer ,
thermometer and reflux condenser .
The reaction mixture was heated at reflux with stirring , for a
period of about one-half hour .
After this time the benzene and unreacted thionyl chloride
were stripped from the reaction mixture under reduced
pressure to yield the desired product N-(β-chloroethyl)-N-
methyl-N'-(2-trifluoromethyl-1,3,4-thiaidazol-5-yl)urea as a
solid residue
90. Text Mining
The N-(β-hydroxyethyl)-N-methyl-N'-(2-trifluoromethyl-1,3,4-
thiadiazol-5-yl)urea prepared in Example 6 , thionyl chloride
( 5 ml ) and benzene ( 50 ml ) were charged into a glass
reaction vessel equipped with a mechanical stirrer ,
thermometer and reflux condenser .
The reaction mixture was heated at reflux with stirring , for a
period of about one-half hour .
After this time the benzene and unreacted thionyl chloride
were stripped from the reaction mixture under reduced
pressure to yield the desired product N-(β-chloroethyl)-N-
methyl-N'-(2-trifluoromethyl-1,3,4-thiaidazol-5-yl)urea as a
solid residue
101. How is DERA going?
• We have text-mined all 21st century articles…
>100k articles from 2000-2013
• Marked up with XML and published onto the
HTML forms of the articles
• Required multiple iterations based on
dictionaries, markup, text mining iterations
• New visualization tools in development – not
just chemical names. Add chemical and
biomedical terms markup also!
108. • The ‘National Compound Collection’
• Extracting compounds manually from theses
• 700 theses, 44,000 compounds (growing…)
• 4 months, 12 UK institutions
• Deposited into ChemSpider
A pilot examining theses
109. • Screening for interesting drug candidates
• Mapping the chain from author to institution
to data to industry
• British Library involved (EThOS collection)
• Build a business model for this
Pilot objectives
110. • Funders encouraging submission from new
dissertations
• Mining of old collections (mostly automated,
likely to need manual QA)
• Extension to other areas of chemical science
…and future (ideal)
111. • Don’t reinvent the wheel
• Research your domain to find work already
underway and test tools for value/utility
In your domain???
Most Domains are Active
Can easily spot titles, authors, abstract, headings
In this case, the OCR used by the scanning organisation couldn’t cope with inderlined titles, and the context pointers were ost
This should be a single node, a noun. But because of the punctuation in the chemical name, the Stanford parser, which is very good but trained on the Wall Street Journal, has interpreted it as a huge baroque sentence.
Not sure who the audience for this will be but worth mentioning if there’s a humanities audience.