=Paper=
{{Paper
|id=Vol-1327/12
|storemode=property
|title=Knowledge Representation of Protein PTMs and Complexes in the Protein Ontology: Application to Multi-Faceted Disease Analysis
|pdfUrl=https://ceur-ws.org/Vol-1327/icbo2014_paper_32.pdf
|volume=Vol-1327
|dblpUrl=https://dblp.org/rec/conf/icbo/RossTLDCCANW14
}}
==Knowledge Representation of Protein PTMs and Complexes in the Protein Ontology: Application to Multi-Faceted Disease Analysis==
https://ceur-ws.org/Vol-1327/icbo2014_paper_32.pdf
ICBO 2014 Proceedings
Knowledge Representation of Protein PTMs and
Complexes in the Protein Ontology: Application to
Multi-Faceted Disease Analysis
Karen E. Ross1, Catalina O. Tudor1, Gang Li1, Ruoyao Ding1, Irem Celen1, Julie Cowart1, Cecilia N. Arighi1, Darren
A. Natale2 and Cathy H. Wu1,2
1
Center for Bioinformatics and Computational Biology, University of Delaware, Newark, DE, USA
2
Protein Information Resource, Georgetown University Medical Center, Washington, DC, USA
E-mail: ross@dbi.udel.edu
Abstract—Alterations in protein post-translational have global effects on the balance of PTM forms in the cell,
modification (PTM) and PTM cross-talk are increasingly being have been implicated in many diseases [1, 2]. Protein
appreciated as driving mechanisms of human disease. The phosphorylation, in particular, has been recognized as a central
Protein Ontology (PRO) is a valuable resource to study the disease-driving mechanism, leading to the development of
relation between PTM and disease because it represents kinase inhibitors as therapeutic agents [1]. With state-of-the-art
individual proteins and protein complex subunits at the text mining tools, it is now possible to extract detailed
proteoform (e.g., isoform, PTM form, and sequence variant) information about proteins, PTMs, and diseases from the
level, with links to their functional properties. We constructed a literature on a large scale. Tools such as RLIMS-P [3] and eFIP
multi-relation network that represents knowledge obtained from
[4] have captured a wealth of information on phosphorylated
large scale text-mining for phosphorylation-dependent protein-
protein interactions (PPIs) and their disease associations, built on
protein forms, their modifying enzymes, and the functional
the PRO framework for representation of PTM forms, impact of phosphorylation, with a minimum of manual curator
complexes, and protein families, as well as their attributes and effort.
relationships. We then conducted two case studies that Despite these advances, representing this information in a
demonstrate the use of PRO in disease analysis. (i) We performed form that is useful for both human interpretation and
cross-species comparisons of two glioma-associated computational reasoning is challenging. It is important not only
phosphorylated proteoforms of the human DNMT1 methylase,
to link genetic variant information with its effects on protein
which revealed that the forms are not strictly conserved in
sequence and function but also to capture the impact of
mouse, a frequently used glioma model system. (ii) We used
PRO-defined proteoforms of the oncoprotein beta-catenin
imbalances of particular PTM forms and complexes on disease.
phosphorylated on various combinations of the N-terminal sites, Used in conjunction with other bioinformatic resources, the
Ser-33, Ser-37, Thr-41, and Ser-45, to interpret a hierarchical Protein Ontology (http://pir.georgetown.edu/pro/pro.shtml;
clustering analysis of cancer types based on their pattern of PRO [5]) enables the structured representation and
mutations in these sites. The cancers formed two major clusters: interpretation of this information. PRO, a member of the Open
one with mutations in Ser-33/Ser-37/Thr-41 and the other with Biomedical Ontologies (OBO) foundry, represents proteoforms
mutations in Thr-41/Ser-45. Proteoform-specific annotation in
(e.g., isoforms, PTM forms, and sequence variants) [6] and
PRO suggests that stabilization of beta-catenin may play a role in
oncogenesis in the first group, whereas alterations in beta-catenin
protein complexes and their relationships within and across
transcriptional or cell adhesion activity may play a more species. Once a proteoform or complex is defined in PRO it
important role in the second group. Together, these scenarios can be annotated with functional and/or disease information
illustrate the general applicability of PRO to disease derived from the scientific literature or bioinformatic
understanding. Future plans include the integration of PRO with databases. This framework can then support the analysis of
other semantic resources to increase our ability to address these biological processes in health and disease.
problems with computational reasoning.
Mouse models have been critical for understanding human
diseases. These models rely on the high degree of conservation
Keywords—Protein Ontology; phosphorylation; text mining;
beta-catenin; cancer of proteins and pathways between human and mouse. Although
protein phosphorylation sites in human are also highly
conserved in mouse (92% conserved in one study [7]),
I. INTRODUCTION conservation at the proteoform level, which can have
Aberrations in protein post-translational modification significant functional consequences in a disease model has not
(PTM), resulting from genetic variations that affect individual been assessed. With its emphasis on representation of
PTM sites as well as alterations in PTM enzyme activity that
43
ICBO 2014 Proceedings
proteoforms and cross-species relationships, PRO is a valuable results, including phospho-dependent PPIs, PTM enzyme-PTM
resource for this type of analysis. form relationships, proteoform/complex-disease relationships,
and relations among PRO terms, is shown in Fig. 1.
In this article, we leverage PRO to: (i) create a
phosphorylation network based on large-scale text mining of This network illustrates several ways in which the PRO
phosphorylation-dependent PPIs that impact disease; (ii) framework allows curators to precisely represent complex
facilitate cross-species comparison of proteoforms of the biological information. First, because PRO treats each
DNMT1 methylase that are associated with glioblastoma in proteoform as a separate entity, multiple forms of a protein can
humans; and (iii) interpret patterns of mutations in the beta- be defined and individually annotated. For example, the Tyr-
catenin oncoprotein observed in different cancer types. These 357-phosphorylated form (PR:000037508) and the Ser-127
examples illustrate the variety of ways in which PRO can be phosphorylated form (PR:000037510) of YAP1 differ in their
used to represent disease knowledge and gain insight into ability to bind 14-3-3 proteins (PR:000003237) and the
disease mechanisms. apoptosis regulatory protein p73 (PR:O15350) and exhibit
different associations with cancer and Alzheimer’s Disease.
II. METHODS PRO can also represent forms with multiple types of
modification, facilitating the description of PTM cross-talk
To identify phosphorylation-dependent PPIs described in (e.g., CCND1 Thr-286-phosphorylated and ubiquitinated form
literature, all PubMed abstracts and PubMedCentral (PMC) (PR:000037512)). Second, PRO represents protein complexes
Open Access full-length articles were processed using eFIP, as distinct entities to which complex-specific annotation can be
which identifies mentions of phosphorylation-dependent PPIs. attached. Moreover, complex subunits are defined using PRO
The kinases, phosphorylated proteins (substrates), and terms, enabling the specification of which particular
interacting partners (interactants) were mapped to UniProtKB proteoforms (e.g., phosphorylated forms) are part of the
identifiers, when possible, using GeneNorm [8]. Disease complex. For example, the proteoform of the DNMT1 DNA
mentions in article titles or abstracts were computationally methlyase that lacks phosphorylation on Ser-127 and Ser-143
detected by matching to a custom dictionary of disease terms (PR:000037504) forms a complex with the DNA-associated
based on MeSH and MedlinePlus. PRO terms for proteoforms factors PCNA (PR:P12004) and UHRF1 (PR:Q96T88). This
and complexes were created as described in [9]. Term complex (PR:000037517) has been associated with tumor
annotation such as binding partners and disease association is suppression [14]. Third, protein terms in PRO are defined at
stored in the PRO annotation file (PAF), and PTM enzyme multiple levels of granularity from the family level down to the
information is recorded in the comments section of the OBO isoform and/or modification level. Thus, when describing a
stanza using structured vocabulary. All terms and annotations biological relationship involving a protein, the term that is
are available upon request and will be made public on the PRO most appropriate given the current state of knowledge can be
website and in downloadable files as part of PRO release 43 used. For example, because 14-3-3 proteins are encoded by
(September 2014). The network was constructed using several genes, and the protein products of these genes are not
Cytoscape 3.0 [10]. Sequence alignment was performed using always distinguishable in experimental assays, 14-3-3 proteins
Clustal Omega [11] and visualized with Jalview 2.8.1 [12]. are represented by the class PR:000003237 that encompasses
Cancer-associated mutations in beta-catenin were obtained the protein products of all 14-3-3 family genes. Similarly,
from Catalog of Somatic Mutations in Cancer (COSMIC) [13]. when the protein is known to be the product of a particular
Tumor tissue types that had at least 10 mutations in the beta- gene, but no isoform information is available, a gene-level
catenin N-terminal phosphorylation sites Ser-33, Ser-37, Thr- PRO term that encompasses all protein products of a gene is
41, and Ser-45 were used. For each tissue the proportion of used (e.g., TP73 (PR:O15350)). Integration of PRO terms into
mutations at each site was calculated relative to the total a multi-relation network context further allows identification of
number of mutations at all four sites. The heatmap was proteoforms sharing common PTM enzymes (e.g., AKT
constructed using the heatmap.2 function of R (version 3.0.2; (PR:000029189)) or interacting partners (e.g., 14-3-3
http://www.r-project.org/) with default parameters. (PR:000003237)) or implicated in the same diseases.
III. RESULTS AND DISCUSSION B. Cross-Species Comparison of Proteoforms of the Glioma-
Associated DNA Methylase, DNMT1
A. Network Representation of Disease-Associated
DNMT1 phosphorylation on Ser-127 or Ser-127/Ser-143 and
Phoshorylation-Dependent Protein-Protein Interactions the concomitant reduction in binding to UHRF1 and PCNA
Text-mining of over 23 million PubMed abstracts and (Fig. 1) has been associated with glioma in humans [14].
800,000 PMC open access full-length articles using eFIP Because the mouse is often used as a model system for
identified sentences describing PPIs that were dependent on the studying glioma, we investigated whether the glioma-
phosphorylation state of one of the interactants in over 13,000 associated DNMT1 proteoforms are conserved in mouse as
articles. About 500 articles also had phosphorylation site well as in several other mammals (Fig. 2). PRO representation
information, UniProtKB-mapped substrates and interactants, of proteoforms enables the cross-species comparison of PTM
and mention of disease in the title or abstract. Through manual at the PTM-form level, which is more likely to reflect
curation of the 109 most recent articles, we found 52 disease- functional conservation than comparisons of the individual
associated phospho-dependent PPIs in 39 articles. (In the sites alone. While human Ser-143 is conserved across all
remaining articles the disease mention was not causally related species, Ser-127 is found only in other primates (red/pink
to the PPI.) A multi-relation network based on some of these
This work has been supported by the National Science Foundation [ABI-
1062520] and the National Institutes of Health [5R01GM080646-08].
44
ICBO 2014 Proceedings
Fig. 1. Multi-relation network showing partial text mining results for disease-associated phosphorylation-dependent protein-protein interactions. PRO terms for
proteoforms and complexes and Disease Ontology terms for diseases are indicated.
residues). Thus, neither the Ser-127 phosphorylated form nor
the Ser-127/Ser-143 phosphorylated form of DNMT1 (Fig. 1) C. Analysis of Cancer-Associated Mutations in beta-catenin
is strictly conserved in mouse, rat, dog, or cow, suggesting that Phosphorylation Sites
non-primates might use a different mechanism for regulating Beta-catenin is a multi-functional protein involved in cell-
DNMT1 interaction with PCNA and UHRF1. However, cell adhesion and transcriptional regulation [16]. Several of its
mouse, rat, and dog (but not cow) have a serine at the adjacent key transcriptional targets drive cell proliferation, and
position (blue/light blue residues), which could potentially excessive beta-catenin transcriptional activity is oncogenic.
fulfill the same role as human Ser-127. This serine has been Beta-catenin stability is regulated by phosphorylation of four
shown to be phosphorylated in mouse in a high-throughput residues in the N-terminus. Casein kinase I phosphorylates Ser-
phospho-proteomic study [15]. Further studies are needed to 45 of beta-catenin, which promotes the sequential
clarify the role of DNMT1 phosphorylation in a mouse glioma phosphorylation of Thr-41, Ser-37, and Ser-33 by GSK3-beta.
model system. Phosphorylation at Ser-37 and Ser-33 enables recognition of
beta-catenin by the ubiquitin ligase beta-TrCP, which targets it
for degradation by the proteosome. Mutations in the four
phosphorylation sites stabilize the protein and have been
associated with cancer.
Through text mining with RLIMS-P and eFIP, we defined
four beta-catenin proteoforms phosphorylated at different
combinations of the N-terminal sites, with distinct sub-cellular
localizations, binding partners, and activities (Fig. 3A). To gain
insight into the role of these proteoforms in cancer, we used
data from COSMIC on cancer-associated mutations in these
sites to perform hierarchical clustering of different cancer types
(Fig. 3B). The cancers fall into two major clusters with
different mutation patterns. Cluster 1 cancers (pink box) are
characterized by mutations at Ser-33 and Ser-37, with few
mutations at Ser-45. Conversely, Cluster 2 cancers (blue box)
are predominantly mutated at Ser-45. Both clusters show
intermediate levels of Thr-41 mutations. The Ser-33/Ser-37
Fig. 2. Partial sequence alignment of the DNMT1 DNA methylase from mutation the pattern in Cluster 1 suggests that oncogenesis in
several mammalian species showing degree of conservation of the human these cancer types is related to the lack of proteoforms 1 and 2.
phosphorylation sites Ser-127 and Ser-143. Both of these forms are unstable due to their association with
45
ICBO 2014 Proceedings
Fig. 3. (A) Proteoforms of beta-catenin phosphorylated on several combinations of the N-terminal phosphorylation sites Ser-33, Ser-37, Thr-41, and Ser-45
with partial functional annotation. (B) Hierarchical clustering of cancer types based on their pattern of mutations in these phosphorylation sites.
[3] X. Yuan, et al., "An online literature mining tool for protein
phosphorylation," Bioinformatics, vol. 22, pp. 1668-1669, 2006.
the ubiquitin ligase beta-TrCP. Thus, beta-catenin stabilization [4] C.O. Tudor, et al., "The eFIP system for text mining of protein
may be playing an important role in these cancers. Kinases, interaction networks of phosphorylated proteins," Database (Oxford),
such as HIPK2, that can phosphorylate Ser-33 and Ser-37 vol. 2012, pp. bas044, 2012.
without prior phosphorylation of Ser-45 may be regulating [5] D.A. Natale, et al., "Protein Ontology: a controlled structured network
beta-catenin stability in these tissues [17]. Cluster 2 cancers of protein entities," Nucleic Acids Res, vol. 42, pp. D415-421, 2014.
have relatively few mutations in the residues that bind beta- [6] L.M. Smith, N.L. Kelleher, and P. Consortium for Top Down,
"Proteoform: a single term describing protein complexity," Nat
TrCP; instead, these cancers are associated with lack of Ser-45 Methods, vol. 10, pp. 186-187, 2013.
phosphorylated proteoform 3. Unlike other beta-catenin
[7] R. Malik, E.A. Nigg, and R. Korner, "Comparative conservation
proteoforms, proteoform 3 is found in the nucleus and may be analysis of the human mitotic phosphoproteome," Bioinformatics, vol.
a key transcriptionally active form of beta-catenin [18]; this 24, pp. 1426-1432, 2008.
proteoform can also bind to the adhesion molecule E-cadherin [8] C.H. Wei and H.Y. Kao, "Cross-species gene normalization by species
[19]. Thus, alterations in beta-catenin transcriptional and cell inference," BMC Bioinformatics, vol. 12 Suppl 8, pp. S5, 2011.
adhesion activity independent of beta-catenin levels may [9] K.E. Ross, et al., "Construction of protein phosphorylation networks by
contribute to Cluster 2 cancers. This example highlights the data mining, text mining and ontology integration: analysis of the
value of integrating experimental disease information from spindle checkpoint," Database (Oxford), vol. 2013, pp. bat038, 2013.
bioinformatic resources such as COSMIC with PRO [10] M.E. Smoot, et al., "Cytoscape 2.8: new features for data integration and
network visualization," Bioinformatics, vol. 27, pp. 431-432, 2011.
representation of proteoforms to gain new insight into disease.
[11] F. Sievers and D.G. Higgins, "Clustal Omega, accurate alignment of
very large numbers of sequences," Methods Mol Biol, vol. 1079, pp.
IV. CONCLUSIONS AND FUTURE WORK 105-116, 2014.
[12] A.M. Waterhouse, et al., "Jalview Version 2--a multiple sequence
Through the structured representation of proteoforms and
alignment editor and analysis workbench," Bioinformatics, vol. 25, pp.
complexes PRO facilitates: (i) representation of proteoform- 1189-1191, 2009.
disease relations identified by large-scale text mining; (ii) [13] S.A. Forbes, et al., "COSMIC: mining complete cancer genomes in the
cross-species comparisons at the proteoform level for Catalogue of Somatic Mutations in Cancer," Nucleic Acids Res, vol. 39,
evaluation of the relevance of animal models of disease; and pp. D945-950, 2011.
(iii) interpretation of disease-associated mutation patterns. [14] E. Hervouet, et al., "Disruption of Dnmt1/PCNA/UHRF1 interactions
Currently, the PRO terms curated in this project can be viewed promotes tumorigenesis from human and mice glial cells," PLoS One,
on the PRO website by biologists interested in PTM, PPI, and vol. 5, pp. e11333, 2010.
disease relationships. We are working toward formalizing these [15] M. Trost, et al., "The phagosomal proteome in interferon-gamma-
activated macrophages," Immunity, vol. 30, pp. 143-154, 2009.
relationships and disseminating them in standard semantic web
[16] T. Valenta, G. Hausmann, and K. Basler, "The many faces and functions
format (e.g. RDF/XML) to enable computational reasoning and of beta-catenin," EMBO J, vol. 31, pp. 2714-2736, 2012.
hypothesis generation.
[17] E.A. Kim, et al., "Homeodomain-interacting protein kinase 2 (HIPK2)
targets beta-catenin for phosphorylation and proteasomal degradation,"
REFERENCES Biochem Biophys Res Commun, vol. 394, pp. 966-971, 2010.
[18] M.T. Maher, et al., "Beta-catenin phosphorylated at serine 45 is spatially
[1] L.M. Graves, J.S. Duncan, M.C. Whittle, and G.L. Johnson, "The
uncoupled from beta-catenin phosphorylated in the GSK3 domain:
dynamic nature of the kinome," Biochem J, vol. 450, pp. 1-8, 2013.
implications for signaling," PLoS One, vol. 5, pp. e10184, 2010.
[2] B.M. Kessler, "Ubiquitin - omics reveals novel networks and
[19] M.C. Faux, et al., "Independent interactions of phosphorylated beta-
associations with human disease," Curr Opin Chem Biol, vol. 17, pp. 59-
catenin with E-cadherin at cell-cell contacts and APC at cell
65, 2013.
protrusions," PLoS One, vol. 5, pp. e14127, 2010.
46