MITRE's Position Paper for W3C Workshop on Semantic Web for Life Sciences

Positional Paper on a Semantic Web for Life Sciences
Alexander A. Morgan, Alexander S. Yeh, Marc Colosimo, Lynette Hirschman,
MITRE Corporation
Contact: amorgan@mitre.org

Our research primarily involves the application of natural language 
processing technology to biomedical literature in support of such 
applications as semi-automated functional annotation of proteins and 
genes, and gene name normalization for improved search and retrieval of 
text information.  We have performed studies in the use of existing 
database resources in these efforts (Morgan, Hirschman et al. 2003) and 
together with CNB/CSIC-Madrid, we have organized and administered a 
challenge evaluation, BioCreAtIvE  (Valencia, Blaschke et al. 2004), 
for text mining systems applied to biomedical literature.  Our primary 
experience with ontologies is with GO (The Gene Ontology Consortium 
2000), and with some of the specific hierarchical controlled 
vocabularies. These include the FlyBase Controlled Vocabulary (The 
FlyBase Consortium 1993) and the TVFac Hierarchy 
(http://www.tvfac.lanl.gov/right.html); our focus has been on automating 
the association of small excerpts of text and the underlying entities 
described (mentioned) in the text with concepts in the ontologies.  We 
focus here on how existing ontologies and related resources can be 
augmented to aid text-mining and how text mining evaluation techniques 
can contribute to ontology evolution, by viewing ontologies as 
annotation guidelines when constructing/populating them.

Even as simple a task as determining which DNA sequence is being 
described when a gene is mentioned in a MEDLINE abstract can be 
challenging task.  With an organism such as Drosophila melanogaster 
(with somewhat free-wheeling naming conventions), identifying the 
mentions of gene names can be non-trivial, given that white, clock, 
dorsal, and period are all gene names.   Trying to associate a gene 
mention with a given functional code in GO is even more difficult, given 
the linguistic distance between a GO concept or description, and how 
this attribute is actually described in text (see the examples given 
below).  We believe that these tasks can be facilitated by enriching the 
lexicon and using sets of synonyms from additional biological resources. 
   If we can automate this process, this will make it far easier to link 
databases and to annotate records in the databases.

An ontology can be enriched by including synonymous descriptions of the 
concept that the node is intended to represent.  A node in a generic 
biological ontology may include a unique identifier, a name, links out 
to parents and children, and sometimes a few sentences describing the 
concept.  However, the name of the concept may be very different than 
anything that might appear in any text mentioning that concept or 
describing an object with its properties.  A longer text description 
might or might not be helpful for these purposes.  For example, an 
ontology of experimental techniques might include a concept such as 
Immunoblot, as does FlyBase.  However, that term is unlikely to appear 
explicitly in the Material and Methods section; rather, we are more 
likely to see the descriptions of antibodies used and a description of 
the procedure.  Another example is the GO code 0005388, 
calcium-transporting ATPase activity, which is unlikely to appear in a 
description of a protein associated with that code.  However, GO 
includes synonyms such as calcium efflux ATPase, calcium pump, and 
sarcoplasmic reticulum ATPase that might aid in recognition, 
particularly combined with a term dictionary that expands calcium to 
Ca2+, the way it most often appears in text.

The effort to develop a large number of LSID’s (Life Science Identifier) 
should be a great help to text-mining efforts.  Linking a GenBank 
accession number with another database with gene and protein annotations 
can help expand the synonymous variants and other key text fields that 
may be used.  Also, mappings between ontologies can help deal with many 
of the previously mentioned issues, because short text descriptions in 
one ontology may be expanded in another.

Unfortunately, research is only just beginning on how to use the links 
between concepts in an ontology to improve text mining.  Taking GO as an 
example, the correct annotation for a protein is the most specific 
(deep) functional annotation known.  Although high up in the ontology, 
concept 0009987, cellular process, exists, and it would be accurate to 
annotate most proteins with this label, it is far too general to be 
relevant.  Computer scientists have examined various distance measures, 
but an underlying problem is that graph distance may bear little 
relation to semantic distance in a human generated ontology.  This sense 
of distance is important when trying to expand a match of terms or 
disambiguate the sense of text in a passage as it relates to an entry in 
the ontology.

The semantic distance is really encoded in how the ontology is used. 
Biological ontologies are used to label data, e.g. associate a GO code 
with a protein, label a patient record with an ICD-10 code, label a 
piece of data with an experimental method code, annotate the subject 
matter of a figure or graph, or link expression of a protein to a 
anatomical term in the FlyBase fly anatomy.  When used in this way, 
ontologies tend to have a very skewed distribution of the labeling. 
(Lord et al 2003) proposes an information theoretic metric based on a 
posteriori distribution of genes annotated in the Gene Ontology.  A set 
of labels with the highest information content would have a uniform 
distribution of the labels.  Instead we see exponential decay curves, 
with a handful of concepts used repeatedly to label different instances, 
and many not used at all.  Repeatedly used concepts may benefit from 
further refinement, e.g. added child concepts to provide more detailed 
information.  The areas of the graph not visited at all show may show 
excess specificity. Of course, the skew also reflects an uneven advance 
of the state of knowledge, as well as trends in research, where certain 
areas receive more attention than others.

When developing an ontology, it is important to keep in mind that it is 
not only a representation of concepts and their relationships, but that 
it generally has specific uses.  In the case of a biological ontology 
that is used to annotate biological ‘entities’, the concepts are 
directly associated with those entities, and it is important to make 
sure that the ontology can be used consistently by different annotators. 
   If two individuals cannot consistently label the same entity with the 
same concepts from the ontology, then there is a problem with how the 
ontology is defined.  These types of inter-annotator experiments that 
use the ontology as annotation guidelines are just now starting to be 
reported in the literature (Camon, Barrell, et al. 2004).

CONCLUSION:

The highly structured nature of ontologies and the semantics they 
represent will provide a valuable resource in natural language 
processing research in the foreseeable future.  Text mining will be 
supported by enriching the text features of ontologies, improving 
indexing for search and retrieval and improving automatic mapping of 
objects to concepts.  Life science ontologies themselves depend on the 
underlying text, since the biological concepts they represent are linked 
to the dynamic literature from which they are drawn.  This would allow 
improved text mining to support efforts to automatically populate 
ontologies.  The natural language processing community can also aid 
ontology design with experience in evaluation and inter-annotator 
studies to create semantic representations of greater utility to both 
human users and automatic systems.

REFERENCES

Morgan, A., L. Hirschman, et al. (2003). Gene Name Extraction Using 
FlyBase Resources. Proceedings of the ACL 2003 Workshop on Natural 
Language Processing in Biomedicine, Sapporo, Japan.
	
The FlyBase Consortium (1993). "FlyBase." GNome News(13): 19-20, 
http://www.geneontology.org.
	
The Gene Ontology Consortium (2000). "Gene Ontology: tool for the 
unification of biology." Nature Genetics(25): 25-29.
	
Valencia, A., C. Blaschke, et al. (2004). BioCreAtIvE Workshop Homepage. 
http://www.pdg.cnb.uam.es/BioLINK/workshop_BioCreative_04/.

P.W.Lord, R.D. Stevens, A. Brass, and C.A.Goble. Semantic Similarity 
Measures as Tools for Exploring the Gene Ontology. In 8th Pacific 
Symposium on Biocomputing (PSB), pages 601-612, 2003.
	
E.B. Camon, D.G. Barrell, E.C. Dimmer, V. Lee, M. Magrane, J. Maslen, D. 
Binns, R. Apweiler. “An evaluation of GO annotation retrieval for 
BioCreative and GOA”.  Journal of Biomedical Informatics, to be published.