Understanding the dynamics of protein-protein interactions (PPIs) is a cardinal step for studying human diseases at the molecular level. Advances in \sequencing" technologies have resulted in a deluge of biological data related to gene and protein expression, yet our knowledge of PPI networks is far from complete. The lack of an integrated vocabulary makes querying this data di cult for domain users, whereas the large volume makes it di cult for intuitive exploration. In this paper we employ Linked Data technologies to develop a framework `LinkedPPI' to facilitate domain researchers in integrative PPI discovery. We demonstrate the semantic integration of various data sources pertaining to biological interactions, expression and functions using a domain-speci c model. We deploy a platform which enables search and aggregative visualization in real-time. We nally showcase three user scenarios to depict how our approach can help identify potential interactions between proteins, domains and genomic segments.
1.1
The study of biological networks forms the integral core of biomedical research related to human diseases and drug development. The ultimate goal of such studies is to understand the connections between di erent genes and proteins, how the cell signals propagate across these networks and regulate their functionality. Hence understanding the Protein-Protein Interaction (PPI) networks underlying each such cellular mechanism is important to speci cally target the dysfunctional proteins, leading towards the discovery of potential drugs and treatments for diseases. Studying PPI networks helps understand the interconnectedness between di erent cellular mechanisms and pathways. Biological pathways are not independent of each other, but their interactions are harmonious, which makes them part of a bigger network. Thus it is important to investigate the dynamics of the cell system as a whole.
Human genome contains more than 20000 protein-coding genes which interact tightly in order to regulate various cellular pathways and mechanisms. The major challenge in developing a thorough understanding of these cellular mechanisms and pathways is to complete the PPI network for each mechanism. However, experimental validation of the binary interactions between total number of proteins is an inconceivable task thus computational models can be used to aid the researchers. These models help identify the sequential, structural and physicochemical properties of known interacting protein pairs and highlight the underlying patterns. Researchers then apply these patterns to narrow down the potential interacting partners for any protein(s) under investigation. Therefore wet-lab validation of the hypothesis formed around the predicted links and protein partners is realistic and achievable.
In computational models, experimentally validated PPIs form the backbone
of PPI networks, however data pertaining to gene-expression, domain-domain
interactions and genomic locations have proved their valuable contribution in
inference and prediction of new links between protein pairs [
Due to advances in sequencing technologies, enormous amount of
experimental data has been generated and stored as independent databases. Databases
such as BioGRID [
The main motivation of this work is to provide researchers with a framework
which enables them to retrieve the answers to their research questions from these
disparate data sources. A researcher interested in the list of protein domains in
a speci c protein can look up the UniProt website3 which is extensively rich
in protein information. Genomic locations of protein-coding genes are publicly
available from several websites such as CellBase [
The challenges in the aggregation and exploration of the aforementioned massive biological data sources have sparked the interests of several domain researchers and led them towards the adoption of a new generation of integrative
technologies, based on Semantic Web Technologies and Linked Data concepts,
thus giving birth to Integrative Bioinformatics [
The nal goal of this research is the identi cation and extraction of potential PPI
networks from various publicly available data sources. The core structure of the
PPI network consists of proteins and their experimentally proven interactions.
Fig. 1 depicts an overview of the LinkedPPI architecture. Following subsections
will describe data selection, RDFization and integration methodologies used.
Validated Interactions: Experimentally validated interactions were retrieved
from BioGRID (Biological General Repository for Interaction Datasets), one
of the most comprehensive PPI databases [
PPI networks and functionality of individual proteins can bene t from identi
cation of essential protein complexes, since di erent subunits contribute to drive
a cellular function. In this work, we have used the latest release of CORUM
(Comprehensive Resource of Mammalian protein complexes) [
One of the crucial challenges in integrative bioinformatics is the heterogeneous nature of biological data sources. Even though several attempts have been made
in the standardization of the data through controlled vocabularies and guidelines,
various hurdles still need to be surpassed. The proteomic standards
initiativemolecular interaction (PSI-MI) is widely accepted by the community for the
modelling of biological networks [
detects!
Experiment! ! "System! "Type! "Scale! ! !
Protein! "UniprotID! !
Gene! "EntrezID! !
hasDomain! Protein!Domain! "PFamID! !
left"right! DD!Interaction! "PFamID! "!Score! "Inferred!Score! partOf! hasGene!
Interaction! "ID! "Source!Database! "Score! "Modification! "Qualification! "Tags! ! !
left"right!
Component! "Official!Symbols! "Synonyms! "Description! !! ! !
hasComponent!
Complex! "ID! "Name! "Synonyms! "Component!Synonyms! "Function!Comment! "Disease!Comment! "SubUnit!Comment! ! ! Ideogram! ""ISDta!(rcth!rm"Ideogram)! left"right! "Stop! "Gene!Count!
Publication! publishedIn! "Title! "Author! "Year! "Abstracts! ! detected!In!
Organism! "TaxonomyID! "Name! ! detectedIn! publishedIn!
II!Interaction! "Inferred!Score! !
We proposed a simple concise domain model, for the modelling of the PPIs
retrieved from BioGRID, complexes from CORUM, protein domains and
genomic location. A domain-speci c model is bene cial over an extensive
wellconstrued ontology due to the absence of non-domain-speci c concepts (Thing,
Continuant, etc.) and is much smaller and self-contained to address a speci c
problem. Being native to a particular domain (e.g. Protein-protein interactions ),
it serves as an intermediate layer between the user and the underlying data, and
enables intuitive knowledge exploration and discovery [
Open Re ne provides a workbench to clean and transform data and
eventually export it in required format. We used its RDF Extension [
After RDFization, the BioGRID data source contains around 11 million triples (11357231), which establish 634996 number of distinct interactions between 14135 Human proteins. The data source also links out to 38952 unique PubMed publications documenting these PPIs. The CORUM data source consists of 156364 triples, with 2867 distinct complexes. The 3did data source consists of 320690 triples with 6818 distinct protein domains and 61582 validated and inferred domain-domain interactions. We inferred a total of 13493 interactions between 405 ideograms, referenced through 80092 triples. 60676 mappings were instantiated between the genes and 7750 extracted GO child leaves.
As such, relevant information could be retrieved from the SPARQL Endpoint
through the formulation of appropriate queries. However as SPARQL requires
a steep learning curve, the non-technical domain user needs intuitive,
interactive visualization tools, which aggregate this information from the multiple data
sources and summarize it. We devised a PPI Visualization Dashboard7 based
on ReVeaLD (Real-time Visual Explorer and Aggregator of Linked Data) [
6 ftp://ftp.uniprot.org/pub/databases/uniprot/current_release/ knowledgebase/idmapping/README 7 http://srvgal78.deri.ie/linkedppi pre-formulated SPARQL SELECT queries8 targeting the various data sources. As shown in Fig. 3, the PPI network associated with the searched protein (e.g., HES1 entrezgene:3280 ) is rendered in a force-directed layout. The list of entities retrieved from the data sources are represented as circular nodes, with the size of each node directly proportional to the number of associated nodes. The nodes are rendered using di erent colors for the sake of visual di erentiation - Red for Components (Proteins of BioGRID or Genes of COXPRESdb), Blue for CORUM Complexes, Light Brown for 3did Protein Domains. The three categories of GO Child Nodes - Biological Processes, Molecular Functions and Cellular Components are displayed using Green, Yellow and Purple colors.
The interactions between di erent proteins are represented as edges - the color of the edges is directly dependent on whether the associations have been retrieved from BioGRID, COXPRESdb or Co-occurrence Data (Black, Red and Purple respectively). The thickness of Black edges depends on the total number of publications which have experimentally validated the underlying interactions. The thickness of the Red and Purple edges depends on the PCC (Pearson Correlation Coe cient for Gene Expression) and Co-occurrence scores respectively. The Protein nodes, which are present in the same complex, possess interacting domains or have underlying coding genes associated to the same GO terms, are not connected directly to each other by edges. They are all connected using similar colored edges to the respective node (complex, domain or GO term), however there may be instances with experimental interactions or co-expression between the connected entities. The resulting network is hence densely clustered, rather 8 https://gist.github.com/maulikkamdar/a47fbecddecc6ba4b373 than a simplistic radial layout of nodes. Hovering over any node highlights subgraph of the network which only displays the rst-level connected nodes and their relations (Fig. 4), hence allowing any domain user to intuitively deduce answers to simple questions like, `Which protein-encoding genes in the network share the same molecular function and have experimental co-expression?' An information box is also displayed beside the hovered node to show additional information like GO term descriptions, Pfam or PubMed IDs, and PCC scores. Zooming and panning across the visualization is possible using a mouse.
The following subsections describe three di erent scenarios depicted in Fig. 5 that our framework could be employed to facilitate extraction of implicit information which can be used as predictors of novel protein-protein interactions. The relevant SPARQL Queries are documented at http://goo.gl/xesMjR. Use Case 1: Extraction of Potential Protein-Protein Interactions Based on the Domain-Domain Interactions. Proteins carry on their functions through their protein domain(s), is a well-known fact. In this scenario we aim to extract possible PPIs based on the known domain-domain interactions. For the sake of simplicity in this use case we assume we are interested in proteins which contain single domains. A researcher has a protein in mind for which the sequence speci cation and domain composition are known. An interesting question might be, list of potential protein partners for this protein. Using our framework researcher can retrieve the list of protein pairs in which at least one of the proteins contains the same protein domain as the protein under question. Possible outcomes are: a) the protein under investigation itself shows up in the result set which forms the list of its experimentally validated protein partners. However this could be queried from the BioGRID web site directly. b) List of proteins with one single domain. In this case with a naive and straightforward conclusion, the researcher may accept the list. However in most cases further application of GO enrichment or advanced statistical analysis o er a more concise list but these analyses are beyond the scope of this work. c) The query results to a set of proteins consisting of several domains which requires further statistical or domain expert knowledge re nement. Despite the need of further investigation in such cases, the shortlisted hypothetical interaction partners are expected to be brief to save a tremendous amount of time and e ort. The SPARQL Query uses the example of the HES1 (entrezgene: 3280 ) protein. We obtain the list of domains present in HES1 - Hairy Orange (pfam:PF07527 ) and HLH(pfam:PF00010 ), and the list of proteins (e.g. HEY2) which share these domains, or have domain-domain interactions. We then retrieve validated PPIs in which the protein participates (e.g. HEY2-SIRT1). We can also obtain the PubMed publication documenting each PPI.
Protein-protein interactions can be identi ed experimentally through various
types of experiments (e.g: Yeast Two-Hybrid). However it is not possible to
identify the interacting domains between two proteins from same experiments
and it requires a set of di erent experiments and protocols. Often protein
domains act as signature elements and repeatedly interact with each other within
the same organism. Therefore these frequent observations assist in identi cation
of novel domain-domain interactions which is enlightening in identi cation of
latent PPIs. Nevertheless in this work such observations are inferred implicitly
from the validated PPI dataset (BioGRID) and require further statistical signi
cance analysis. In our SPARQL Query example, we retrieve the validated and the
inferred scores for domain interactions with the HLH domain (pfam:PF00010 ).
Use Case 3: Identi cation of Selective Interactions between Segments
of Human Genome. Human chromosomes are compact in 3D space with each
chromosome folding into its own territory [
In this work we have selected the boundary of ideogram bands on each chromosome as genomic location of each gene. Several genes may reside on one ideogram as well as genes that may fall between two consecutive ideograms. The reason for this selection of distance unit is to take into account the e ect of the co-expression of neighboring genes and the possible shared mechanism. Protein pairs from PPI dataset are mapped to their genomic location and simple frequency calculation from retrieved data can identify the signi cance of interactions between two genomic segments. Based on these ndings researchers can propose the genomic location pattern in which proteins preferentially select their interacting partners. These regions may contain genes involved in the same pathways or share the same functionality which are yet to be identi ed by further gene enrichment analysis. As shown in the provided SPARQL Query, we could retrieve the pre-determined inferred score between two ideograms, e.g. Chromosome 3 - q29 and Chromosome 10 - q24.32 containing the protein-coding genes for HES1 and SIRT1 (entrezgene:23411 ) proteins respectively.
Query Protein - > PX PX - PY PD - PC PD - PF PA - PF (a) x x x d d x x y
Query Domain - > Dx
x Dx-Dy inferred Dx-Dc validated Dx-Dw validated Dx-Df inferred Dx-Dy inferred (b) x x x x x y y
PA
(c)
ga
gb
PB
Jiang et al. developed a semantic web base framework which predicts targets of
drug adverse e ect based on the PPIs and gene functional classi cation algorithm
[
The incorporation of complementary datasets for the expansion of PPI networks is a useful approach to gain insight into biological processes and to discover novel PPIs which have not been documented in the current PPI databases. However, there is an inherent high level of heterogeneity at the schema and instance level of these data sources, due to lack of a common representation schema and format. Hence, we decided to apply Linked Data concepts in the integration, retrieval and visualisation of concealed information. The enormous amount of publicly available data and its dynamicity, in terms of regular updates, is currently a
rate-limiting step to our data-warehousing approach for centralised analysis. We have proposed a domain-speci c model which can accommodate the needs in the eld of PPI modelling. The use of a domain-speci c model and an interactive graph-based exploration platform for search and aggregative visualisation makes our integration approach more intuitive for the actual users who deal with PPI predictions. We have also proposed a set of three user scenarios depicting how LinkedPPI framework could be used for the prediction of potential interactions between proteins, domains and genomic regions. 6
The approach which has been presented in this work is used in extraction of valuable information with regard to PPI network, domain-domain interactions and selective genomic interactions. However the observations reported in the outcome of such data retrieval is raw and could be a valuable asset for simulations and prediction methods if further analysis is done. As part of the future work we intend to apply statistical analysis on signi cance of such observations in order to be able to develop a classi er algorithm which is able to predict interacting and non-interacting protein pairs.
Acknowledgements This work has been done under the Simulation Science program at the National University of Ireland, Galway. SimSci is funded by the Higher Education Authority under the program for Research in Third-level Institutions and co-funded under the European Regional Development fund.