Text mining in the field of
evolutionary biology: facilitating
scholarly collaboration
Sarah Carrier
February 2008
What is text mining?
• Deriving novel, relevant information from
unstructured information (text).
• Identification of patterns and trends.
• Typical techniques:
– Clustering
– Categorization
– Concept/entity extraction -> dictionary-based, statistical
methods/machine learning
– Document summarization
Long-Term Objective
1. To identify biological entities through text
mining methods, then categorize them into
predetermined classes of objects
2. To describe biological concepts using
simple ontologies - for example, use the
controlled vocabulary generated in step 1 to
describe results and methods
Semester Objective
1. To categorize evolutionary biology
abstracts into 5 different predetermined
categories using nouns and noun-phrases
associated with the text.
2. To prepare for long-term objectives.
Motivation
• Scholarly collaboration
• Generation of ontologies to describe results
of experiments, to enhance meta-analyses
for research purposes
• Web publishing
• Indexing by central repositories
Motivation and Current Research
• need in the life sciences for alternatives to keywordbased approaches based in the traditional information
retrieval framework
• extensive (text mining) work is being done to identify
protein-protein interactions and gene annotations
• extracted entities can be linked to existing ontologies
and potentially used to generate new ontologies
• the most common text mining applications in the life
sciences tend toward information extraction, as this
method produces a potential solution to the deluge of
information in the field
Manual Keyword Identification
• 8 categories: concept, field/discipline, gene,
habitat, method, place, taxon, time period
• 104 articles, 5 journals, 600 keywords - 551 with
duplicates removed, most terms ended up in the
“concept” category -> varied sizes
• Manual categorization accomplished with domain
experts on the Dryad team, matched with existing
terminologies
• 16% were duplicates, avg. 50% matched
terminologies - implies that controlled
vocabularies should be used for standardization
Some potential challenges
• Evolutionary biology is an interdisciplinary field:
ecology, genomics, paleontology, population
genetics, physiology, systematics
• A varied and complex terminology for the life
sciences
• Incredibly sparse dataset
• Coverage of existing terminologies incomplete
(UMLS, Open Biomedical Ontologies)
Methodology
• MEDLINE abstracts from American Naturalist,
Ecology, Journal of Evolutionary Biology,
Molecular Ecology, Molecular Biology and
Evolution, Systematic Biology
• Total: 15,179 abstracts, 227,731 terms extracted
from list of MeSH terms and 831,245 terms using
abstract
• Standard preprocessing of abstracts using Perl,
including the Porter stemmer and the Brill Tagger
An Example
PMID- 17206577
TI- Ecological specialization and adaptive decay in digital organisms.
AB- The transition from generalist to specialist may entail the loss of
unused traits or abilities, resulting in narrow niche breadth. Here we
examine the process of specialization in digital organisms--selfreplicating computer programs that mutate, adapt, and evolve. Digital
organisms obtain energy by performing computations with numbers
they input from their environment. We examined the evolutionary
trajectory of generalist organisms in an ecologically narrow
environment, where only a single computation yielded energy.
CONTINUED…
MH- *Adaptation, Biological, Competitive Behavior, Computer
Simulation, Ecology, *Evolution, Molecular, Genotype, *Models,
Genetic, Mutation, Phenotype, Software
Preprocessing
17206577|1|transition
17206577|1|specialist
17206577|1|loss of unus trait
17206577|1|trait
17206577|1|generalist
17206577|1|loss
17206577|1|transition from generalist
17206577|1|unus trait
17206577|1|narrow nich breadth
17206577|1|nich breadth
17206577|1|breadth
17206577|2|process
17206577|2|abil
17206577|2|nich
• CONCEPT: regressive
evolution,
specialization,
pleiotropy, adaptation,
mutation accumulation
• METHOD: digital
evolution
Preprocessing, cont.
The/DET transition/NN from/IN generalist/NN to/TO specialist/NN
may/MD entail/VB the/DET loss/NN of/IN unused/JJ traits/NNS
or/CC abilities/NNS ,/PPC resulting/VBG in/IN narrow/JJ niche/NN
breadth/NN ./PP Here/RB we/PRP examine/VBP the/DET process/NN
of/IN specialization/NN in/IN digital/JJ organisms/NNS selfreplicating/NN computer/NN programs/NNS that/IN mutate/VB ,/PPC
adapt/VBP ,/PPC and/CC evolve/VB ./PP Digital/NNP organisms/NNS
obtain/VBP energy/NN by/IN performing/VBG computations/NNS
with/IN numbers/NNS they/PRP input/NN from/IN their/PRPS
environment/NN ./PP We/PRP examined/VBD the/DET
evolutionary/JJ trajectory/NN of/IN generalist/NN organisms/NNS
in/IN an/DET ecologically/RB narrow/JJ environment/NN ,/PPC
where/WRB only/RB a/DET single/JJ computation/NN yielded/VBD
energy/NN ./PP We/PRP determined/VBD the/DET extent/NN to/TO
which/WDT
An Example
• <MeshHeadingList>
• <MeshHeading>
• <DescriptorName MajorTopicYN="N">Adaptation,
Physiological
• </DescriptorName>
• </MeshHeading>
• <QualifierName MajorTopicYN="N">genetics
• </QualifierName>
• <QualifierName MajorTopicYN="Y">metabolism
• </QualifierName>
• </MeshHeading>
• <MeshHeading>
• <DescriptorName MajorTopicYN="N">Predatory Behavior
• </DescriptorName>
• </MeshHeading>
• </MeshHeadingList>
Most Frequent
Abstrac t Terms
(collection)
speci
popul
gene
result
sequenc
studi
data
ana lysi
pattern
evo lut
variat
dna
phylogene t
region
model
leve l
rate
ana lys
structur
select
Most Frequent MeSH
Terms (collection)
gene t
sequenc
anim
dna
phys iologi
phylogen i
evo lut
popul
data
ana lysi
model
molecular sequenc da ta
sequenc da ta
gene
variat
acid
base
base sequenc
classif
protein
Other Steps
• TF*IDF weighting, pruning
– Challenges: skew in category sizes (“concept” being the
largest), lack of truly discriminative terms
• Application of a machine-learning model: Hidden
Markov Models, Support Vector Machines
– SVMs: outperform HMM
• also better for large, sparse datasets
• Evaluation:
– Recall, Precision, F-Scores
– Presentation to Dryad domain experts for feedback
Future Steps
• Use of existing vocabularies to assist in
controlling terminology: NBII thesaurus,
MeSH, GTN, WordNet, Gene Ontology,
ITIS, UBIO, UMLS, etc.
Ontology generation?
• The POS processing has already been done - the
verb is an essential element of the relationship
• Find most common verbs and define them as
“relational verbs”
• Methodology: using POS tags, pull out “triplets”
or certain sequences of words
– NOUN - VERB - NOUN
…in some studies, prepositions are also analyzed
Ontology, cont.
Our/PRPS results/NNS show/VBP that/IN as/IN
organisms/NNS evolved/VBD improved/VBN
performance/NN of/IN the/DET selected/JJ
function/NN ,/PPC they/PRP often/RB lost/VBN
the/DET ability/NN to/TO perform/VB other/JJ
computations/NNS ,/PPC and/CC these/DET
losses/NNS resulted/VBD most/JJS often/RB
from/IN the/DET accumulation/NN of/IN
neutral/JJ and/CC deleterious/JJ mutations/NNS
./PP
Conclusions
• Term variation and ambiguity presented a challenge in
my project because it yielded a very sparse data set
• With more time I would have supplemented the
dataset I generated this semester with more data from
more abstracts, perhaps even the full text, if available
• Although the objective of the project changed over the
semester, the results provide valuable insight into the
structure and use of evolutionary biology vocabularies
• Potential future developments in the project, namely
ontology generation, would have a positive impact on
scholarly communication amongst researchers in the
field of evolutionary biology
Thank you!