=Paper=
{{Paper
|id=Vol-2285/ICBO_2018_paper_52
|storemode=property
|title=The Integrative use of Anatomy Ontology and Protein-protein Interaction Networks to Study Evolutionary Phenotypic Transitions
|pdfUrl=https://ceur-ws.org/Vol-2285/ICBO_2018_paper_52.pdf
|volume=Vol-2285
|authors=Pasan Fernando,Erliang Zeng,Paula Mabee
|dblpUrl=https://dblp.org/rec/conf/icbo/FernandoM18
}}
==The Integrative use of Anatomy Ontology and Protein-protein Interaction Networks to Study Evolutionary Phenotypic Transitions==
https://ceur-ws.org/Vol-2285/ICBO_2018_paper_52.pdf
Proceedings of the 9th International Conference on Biological Ontology (ICBO 2018), Corvallis, Oregon, USA 1
The integrative use of anatomy ontology and
protein-protein interaction networks to study
evolutionary phenotypic transitions
Pasan C. Fernando, Erliang Zeng, Paula M. Mabee
Biology Department
University of South Dakota
Vermillion, USA
pasan.fernando@coyotes.usd.edu, erliang.zeng@usd.edu, paula.mabee@usd.edu
Abstract— Studying evolutionary phenotypic transitions, such interactions that are important in regulating the phenotype.
as the fin to limb transition, is popular in evolutionary biology. Understanding the modular structure of gene interactions is
The recent advances in next-generation technologies have extremely important in studying their role in the development
accumulated large volumes of genomics and proteomics data, of phenotypes because it is the gene interactions that
which can be used to analyze the genetic basis for evolutionary determine the outcome rather than the individual genes.
phenotypic transitions. Protein-protein interaction (PPI)
networks can be used to predict candidate genes and identify The biggest challenge of using PPI networks is the low
gene modules related to evolutionary phenotypes; however, they candidate gene prediction accuracy due to the low quality of
suffer from low gene prediction accuracy. Therefore, an
the networks [1]. The PPI networks are known to contain a
integrative framework was developed using PPI networks and
anatomy ontology, which significantly improved the accuracy of
higher amount of false positive interactions, and some
network-based candidate gene predictions in zebrafish and networks are still incomplete [2]. Before using PPI networks
mouse. This integrative framework will also be used to identify to study evolutionary phenotypic transitions, their quality
gene modules associated with the fin to limb transition and to must be improved to obtain better results. Because we are
study the changes in these modules which lead to the phenotypic focusing on anatomical phenotypes, such as the pectoral fin
change. development and the forelimb development, we propose an
integrative framework that uses anatomy ontology to
Keywords- Anatomy ontology; network analysis; protein- incorporate known information about gene-phenotype
protein interactions; data integration; gene prediction. relationships in literature with the PPI networks. This
integration is expected to improve the PPI network quality and
I. INTRODUCTION predict candidate genes with a higher accuracy. To test this
The process of evolution is accompanied by numerous hypothesis, we use known anatomical phenotype annotations
important phenotypic transitions, such as the fin to limb from mouse and zebrafish. After the evaluation, the integrated
transition in vertebrates, which contributed to the wealth of networks will be used to detect gene modules associated with
phenotypic diversity observed among different species today. the fin to limb transition in mouse and zebrafish, and the
Understanding the relationship between genes and their modules will be compared to observe the genetic changes
phenotypes is important in explaining the changes in those corresponding to the phenotypic transition.
phenotypes. Traditionally, wet lab methods were used to
discover genes to phenotype relations. Despite the higher II. METHODS
accuracy of their predictions, wet lab candidate gene The first step of the integrative framework is constructing
prediction methods are high in resource and time gene networks that are entirely based on the known gene to
consumption, which lead to the popularity of faster anatomical phenotype annotations. The anatomical profiles
computational candidate gene predictions methods [1] that use for mouse and zebrafish were downloaded from the Monarch
the genomic and proteomic data accumulated in public initiative data repository (https://monarchinitiative.org/),
databases. which retrieves data from model organism databases.
Monarch initiative data is manually pre-processed to remove
The use of PPI networks for candidate gene prediction has unwanted annotations and the genes are annotated to Uberon
become popular due to the availability of large PPI datasets anatomy ontology terms [3]. Uberon (http://uberon.github.io/)
for model organisms. Network analysis algorithms can be is a cross-species anatomy ontology that integrates species-
used to analyze PPI networks and detect gene modules specific anatomy ontologies, such as Mouse Anatomy
corresponding to phenotypes in question [1]. Other gene Ontology (MA) and Zebrafish Anatomy Ontology (ZFA),
prediction methods only discover direct gene to phenotype which makes it suitable for evolutionary analyses involving
relationships, but network analysis further identifies gene multiple species [4].
ICBO 2018 August 7-10, 2018 1
Proceedings of the 9th International Conference on Biological Ontology (ICBO 2018), Corvallis, Oregon, USA 2
Semantic similarity scores between anatomy ontology
terms were calculated to obtain pairwise gene similarity
values for all the genes in mouse and zebrafish. Semantic
similarity is a quantitative value that represents similarity
between two ontology terms based on their location in the
ontological structure and their gene annotations [5]. Four
different semantic similarity methods (Lin, Resnik, Schlicker,
and Wang) were used to generate pairwise gene similarity
matrices, which in turn were used to generate gene networks
that are entirely based on the anatomy ontology annotations of
the genes (anatomy-based gene networks). These networks
were filtered using a gene similarity score cutoff to remove
interactions with low scores. In these networks, the genes with
higher similarity scores are the ones that are annotated to
similar anatomy ontology terms.
The PPI networks for mouse and zebrafish were
downloaded from the STRING database (https://string-
db.org/). Then, the PPI networks were integrated with the
anatomy-based gene networks using pairwise gene similarity Fig. 1. Comparison of ROC curves for PPI (red), integrated (green), and
scores of the two networks in a probabilistic model. In the anatomy-based gene (blue) networks for the four semantic similarity
methods. The integrated and anatomy-based gene networks clearly
integrated network, only the gene pairs that receive high
outperform the PPI networks when predicting candidate genes.
similarity scores from both the input networks have high gene
similarity scores. To assess the candidate gene prediction large number of unknown genes coming from PPI networks,
performance of the integrated networks and the PPI networks, which can be potential candidates for anatomical phenotypes.
Uberon anatomy ontology terms that have at least 10 or more Therefore, integrated networks are more useful for
gene annotations were used from the zebrafish and mouse downstream network analysis.
anatomical profiles downloaded from the Monarch initiative
data repository. Hishigaki prediction method [6] was used as
The integrated network with the highest performance
the network-based candidate gene prediction algorithm and for mouse and zebrafish will be used for detecting gene
leave-one-out-cross-validation was used as the evaluation
modules associated with the fin to limb transition. Because the
technique. Receiver operating characteristic (ROC) and quality of the integrated networks is higher than the PPI
precision-recall curves were generated for the comparison of
networks, the gene modules will be more accurate. The gene
different network types. Although the goal was to compare the modules for pectoral fin and pelvic fin in zebrafish will be
integrated versus PPI networks, the anatomy-based gene
compared with gene modules for forelimb and hindlimb in
networks were also included in the comparison. mouse, respectively, to identify modular changes genes during
the fin to limb transition. This work showcases how anatomy
III. PRELIMINARY RESULTS AND DISCUSSION ontology can be used to improve the quality of candidate gene
The ROC and precision-recall curve comparisons for predictions and to perform efficient network analyses to study
mouse and zebrafish indicate that the integrated networks evolutionary transitions.
significantly outperform the original PPI networks when
predicting candidate genes (Only the zebrafish ROC curve REFERENCES
comparisons of the four semantic similarity calculation [1] R. Sharan, I. Ulitsky, and R. Shamir, "Network-based prediction of
methods are shown in Fig. 1). This result is consistent among protein function," Molecular systems biology, vol. 3, 2007, p. 88.
the four semantic similarity calculation methods used. The [2] C. von Mering et al., "Comparative assessment of large-scale data sets
higher candidate gene prediction accuracy of the integrated of protein-protein interactions," Nature, vol. 417, 2002, pp. 399-403.
networks means that their network quality was increased [3] C. J. Mungall et al., "The Monarch Initiative: an integrative data and
during the integration. Although anatomy-based gene analytic platform connecting phenotypes to genotypes across species,"
networks (shown in blue in Fig. 1) have the highest Nucleic Acids Res, vol. 45, 2017, pp. D712-D722.
performance among most of the semantic similarity [4] M. A. Haendel et al., "Unification of multi-species vertebrate anatomy
ontologies for comparative biology in Uberon," Journal of biomedical
calculation methods, they are not suitable for candidate gene semantics, vol. 5, 2014, p. 21.
prediction or identifying network modules because they only
[5] C. Pesquita, D. Faria, A. O. Falcão, P. Lord, and F. M. Couto,
contain genes that have at least one anatomy ontology term "Semantic Similarity in Biomedical Ontologies," PLoS Comput Biol,
annotation. This number is low compared to the integrated and vol. 5, 2009, p. e1000443.
PPI networks. For instance, the zebrafish anatomy-based gene [6] H. Hishigaki, K. Nakai, T. Ono, A. Tanigami, and T. Takagi,
network constructed using the Schlicker method contains "Assessment of prediction accuracy of protein function from protein–
5,386 genes, whereas the corresponding integrated network protein interaction data," Yeast, vol. 18, 2001, pp. 523-531.
contains 12,755 genes. The integrated networks contain a
ICBO 2018 August 7-10, 2018 2