1. Integrating Large, Disparate, Biomedical
Ontologies to Boost Organ Development
Network Connectivity
Chimezie Ogbuji1 and Rong Xu2
Metacognition LLC1
Case Western Reserve University2

2. Outline
 Outline
◦ Background
◦ Motivation
◦ Literature review / related work
◦ Opportunity / specific example
◦ Hypothesis
◦ Method
◦ Evaluation
◦ Discussion

3. Background
 Controlled biomedical vocabulary systems
(and ontologies) play a key role in the
analysis of genetic disease
◦ Structured, interoperable, and machine-readable
◦ Facilitate reproducibility of scientific results and
use of intelligent software that can leverage
underlying meaning
◦ Scientific results and the structured biomedical
knowledge they are based on may be used for
multiple - even unanticipated - purposes

4. Motivation
 Want descriptive relations that comprise
terminology paths between (congenital)
diseases and the anatomical entities that
become malformed
 Want to use these as the basis for
analysis and classification of congenital
disorders according to their underlying
molecular mechanism

6. Opportunity
 The Gene Ontology (GO) is arguably the most prominent
example of how highly-organized and structured medical
knowledge can be leveraged to facilitate medical genetics
◦ Has a hierarchy of biological processes involving organ
development.
 The Foundational Model of Anatomy (FMA) is a vast
ontology with an objective to conceptualize the physical
objects and spaces that constitute the human body
◦ macroscopic, microscopic and sub-cellular canonical anatomy.
 Their skeletal relations (is_a, part_of, and has_part) have the
same meaning

7. Opportunity (continued)
 Their skeletal relations (is_a, part_of, and
has_part) have the same meaning
 There are no immediately usable
terminology paths between concepts in
the GO's anatomy development process
hierarchy and participating anatomical
entities defined in the FMA

8. Literature review
 Cellular components function via interaction with
each other in a highly-complex and
interconnected network
 Interdependencies among a cell’s molecular
components lead to functional, molecular, and
causal relationships among distinct phenotypes.
 Network-based approaches to disease have the
potential to provide a framework for classifying
disease, defining susceptibility, predicting disease
outcome, and identifying tailored therapeutic
strategies
Barabási et al. Network Medicine: A Network-based Approach to Human Disease, Nature Reviews
Genetics 2011.

9. For over a decade, analysis of biological networks via network and graph theory
has revealed the importance of locally-dense and
well-connected subgraphs (hubs).
Schwikowski et al. A network of protein-protein interactions in yeast 2000
Barabási et al. 2011

10. Related work
 Investigation of structural and lexical
concordance between anatomy terms in the FMA
and SNOMED-CT
◦ Bodenreider & Zhang 2006
 Leveraging this concordance for integrating
modules from each for a specific domain
◦ Ogbuji et al. 2010
 Discussion of logical consequences of using
part_of between both anatomical entities (in the
FMA) and biological processes (the GO)
◦ Jimenez-Ruiz et al. 2010

11. Opportunity: Cardiovascular
disease and development
 Understanding the formation of the heart is
critical to the understanding of
cardiovascular diseases
 The study of genes and gene products
involved in cardiovascular development is an
important research area
 There have been recent efforts to expand
the subset of the GO's anatomy
development hierarchy involved in heart
development

12. Marfan Syndrome (MFS)
[…] mainly characterized by aneurysm formation in the
proximal ascending aorta, leading to aortic dissection
or rupture at a young age when left untreated. The
identification of the underlying genetic cause of MFS, namely
mutations in the fibrillin-1 gene (FBN1), has further
enhanced [...] insights into the complex pathophysiology of
aneurysm formation
In UMLS Metathesaurus
• Finding site: connective tissue structure (SNOMED-CT)
• Category: congenitial skeletal disorder (CRISP Thesaurus and NLM MTH)

13. Marfan Syndrome example
 In the GO, FBN1 is annotated with the
GO_0001501 (skeletal system development)
and GO_0007507 (heart development)
concepts (amongst others)
 The former coincides with the more common
finding site and classification of MFS as a
congenital skeletal disorder
 This is in spite of the fact that associations (causal
and otherwise) between MFS and cardiovascular
diseases such as aortic root dilation are well-
documented in the medical literature

14. Hypothesis
 A high-quality integration of the GO's
development process hierarchy with the FMA will
have several benefits:
◦ New biological pathways from genetic diseases to the
anatomical entities whose development are involved
in their underlying molecular mechanisms
◦ Graph and network analysis can benefit from an
increase in connectivity for discovering biologically
meaningful motifs
◦ Similarly, classification algorithms can also take
advantage of this

15. Copper: annotates human gene
Gold : does not annotate human gene

16. Method and materials
 Integration is performed on the following GO
development process hiearchies
◦ Anatomical structure development
◦ Anatomical structure arrangement
◦ Anatomical structure morphogenesis
 Only GO concepts that annotate human genes
are considered
 In processing the GO, the logical properties
(transitivity, for example) of the relations are fully
considered
◦ This will always be the case, henceforth

17. Method and materials (continued)
 The FMA ontology is loaded (as OWL/RDF) into a
triple store for remote querying via SPARQL
 The prefix of the human-readable label for each GO
concept in the development hierarchies is stemmed
and used as a basis for case-insensitive, lexical
matching on primary labels and exact synonyms of
FMA classes via a SPARQL query
 FMA classes that match exactly are considered to
denote the anatomical entities that participate in the
corresponding GO biological process

18. Example
 GO_0007507 (heart development)
 Prefix: heart
 Matching FMA concept: FMA_7088
(Heart)

19. Evaluation
 Result: 1644 development process and
anatomical entity pairs
 We calculate the Jaccard coefficient of
the overlap between hierarchies for 6
major organs and the anatomical
development processes they participate
in

20. Evaluation (continued)
 Using the GO development process for some
FMA organ O as the starting point, the set of all
subordinate terms is calculated: GOsubgraph(O)
 Example:
◦ GO_0007507 (heart development) has
GO_0003170 (heart valve development) as a
component (via has_part)
◦ GO_0003170 subsumes GO_0003176 (aortic valve
development) and has GO_0003179 (heart valve
morphogenesis) as a component
◦ Each of these would be considered as subordinates of
GO_0007507

21. Evaluation (continued)
 In a similar fashion, the subordinate anatomical
entities for each O amongst the 6 chosen organs
are calculated:
◦ FMAsubgraph(O)
 For each O, we calculate the GO terms that are
both in GOsubgraph(O) and were matched with an
FMA class that is in FMAsubgraph(O)
 This resulting set of GO terms is considered the
intersecting set and the Jaccard coefficient is
calculated with respect to this, FMAsubgraph(O), and
GOsubgraph(O)

22. Jaccard Coefficient (overlap)

23. Evaluation: network connectivity
 We calculate number of new paths from
OMIM diseases through their genes to
the anatomical entities in the FMA:
◦ P+dgo
 Similarly, we calculate the number of new
paths starting from the genes to
additional FMA anatomical entities
◦ P+go

24. Network connectivity: continued
 Only genes that are annotated with
anatomical development processes
matched to FMA classes and OMIM
diseases associated with these genes
were considered
◦ Genesdev

25. Number of additional P+dgo paths on a logarithmic scale

26. Histogram of the distribution of additional P+dgo paths as a whole
and normalized by the number of genes associated with each disease

28. Log-scaled histogram of additional paths from Genesdev to FMA
classes, only for those genes that had additional paths

29. Evaluation summary
 On average, mapping introduces 9,549
additional P+dgo paths per OMIM disease
 On average, each Genedev gene had 17,037
additional paths to FMA classes
 Caveat in normalizing the number of P+dgo
paths by number of genes
◦ paths from diseases to anatomical entities
introduce combinatorial factor of disease-gene
pairings

30. Discussion
 Overlap results indicate little overlap
between the GO hierarchies and
corresponding FMA hierarchies
 Not surprising as both cover disparate
domains within medicine and one is
specific to humans while the other is not

31. Discussion (continued)
 This along with the size of the FMA as a
whole and within the portions mapped to
the GO hierarchies indicate opportunity to
build on the mapping and to integrate both
ontologies in a meaningful way
 Connectivity results demonstrate significant
increase of biological paths from genetic
diseases (and their genes) to the anatomical
entities participating in the development
process

32. Discussion (continued)
 As these paths are at least as logically and
biologically sound as the ontologies they
were forged from, we expect that an
appreciable amount of them will be useful
for analysis
 To our knowledge, this is the first attempt
of this kind to integrate the anatomical
structural development, morphogenesis, and
organization hierarchies in the GO with the
FMA

33. Limitations
 Regarding deductions (formal or
otherwise) that follow from an
integration of the FMA and GO
◦ Need to be careful to only consider
annotations for humans or to have a robust
way to manage the uncertainty introduced in
not doing so