Customizable views on semantically integrated networks for systems biology.

MOTIVATION: The rise of high-throughput technologies in the post-genomic era has led to the production of large amounts of biological data. Many of these datasets are freely available on the Internet. Making optimal use of these data is a significant challenge for bioinformaticians. Various strategies for integrating data have been proposed to address this challenge. One of the most promising approaches is the development of semantically rich integrated datasets. Although well suited to computational manipulation, such integrated datasets are typically too large and complex for easy visualization and interactive exploration.
RESULTS: We have created an integrated dataset for Saccharomyces cerevisiae using the semantic data integration tool Ondex, and have developed a view-based visualization technique that allows for concise graphical representations of the integrated data. The technique was implemented in a plug-in for Cytoscape, called OndexView. We used OndexView to investigate telomere maintenance in S. cerevisiae. AVAILABILITY: The Ondex yeast dataset and the OndexView plug-in for Cytoscape are accessible at http://bsu.ncl.ac.uk/ondexview.

1 INTRODUCTION

1.1 Systems biology and integrative bioinformatics

Systems biology considers multiple aspects of an organism's structure and function at the same time, using the plethora of data that is publicly available online. Biologists have access to heterogeneous data covering many different aspects of biology; 1230 entries are listed in the 2010 database issue of Nucleic Acids Research (http://nar.oxfordjournals.org/content/38/suppl_1). For model organisms such as Saccharomyces cerevisiae, the abundance of freely available data has long since exceeded the point at which it can be analysed manually.

Most data sources still exist in isolation, each with its own specialization and focus (Stein, 2002). In many cases, databases lack semantic links to each other, even when they are providing data about the same entities. This partitioning of knowledge is a particular problem for systems biology, hampering the examination of the emergent behaviours inherent in complex biological systems. The problem of using distributed, heterogeneous datasets is addressed by the subdiscipline of integrative bioinformatics. There are many different approaches to the integration and querying of large heterogeneous datasets. Early XML-based approaches (Achard ) were soon replaced by more sophisticated methods. Federated approaches, such as the distributed annotation system (DAS) (Prlić ) establish a central view on the data by translating each internal query into a set of external queries in order to retrieve the required data from the outside sources on the fly (Heimbigner and McLeod, 1985). Semantic web approaches, as endorsed by the Open Biological and Biomedical Ontologies foundry (OBO) (Smith ) or as implemented in YeastHub (Cheung ), establish a network of interlinked ontologies using a set of controlled vocabularies (Brinkley ). Data warehousing approaches such as BioMART (Haider ) EcoCyc (Keseler ) and SAMBA (Tanay ), convert and incorporate data from their external sources into their own database schema and provide custom queries.

1.2 Ondex

Ondex is a graph-based data integration framework (Kohler ), which takes a semantic warehousing approach. Links between entities in different datasets may be directly extracted from the available data sources, inferred as part of the integration, or generated by external tools, such as BLAST (Altschul ). These data can subsequently be matched and interlinked with each other semantically. The result is a semantically enriched dataset, with which users can interact and also visualize.

Ondex incorporates data into a network of entities termed ‘concepts’, connected by ‘relations’, all of which can carry ‘attributes’. All concepts, relations and attributes have ‘types’, which are organized in a hierarchical fashion. For example, the concept type Protein is a subtype of Molecule, which is itself a subtype of Thing. This hierarchy means that every Protein concept is also a Molecule and a Thing. Similarly, the relation type catalyzes is a subtype of actively_participates_in. Therefore, every statement that p::Protein catalyzes r::Reaction means that p actively_participates_in r. Adding this type of information means that the computer stores not only data, but also its meaning, providing a ‘semantic representation’ of the data.

A special kind of attribute that all concepts in the an Ondex graph have is the cross-reference (called ‘concept accession’ in Ondex). Concept accessions make it easy to connect concepts originating from different sources. For example, if an Ondex dataset contains two Gene concepts which have been imported from two different data sources, but which share the same concept accession, Ondex can connect these concepts with a new relation of type same_as, and can at a later time merge these concepts into a single concept. All concepts and relations also have attached provenance information stating their origin and any associated evidence codes.

The process of integrating data into this data structure is performed by the Ondex workflow engine, which employs a plugin architecture. Parsers, mapping methods and other plugins can be developed using an open API (Taubert ) and can subsequently be used in workflows.

In this work, we describe a mechanism to collapse groups of concepts into a simpler, more easily visualized and conceptualized representation. We have implemented this mechanism as a plugin for Cytoscape, called OndexView, which facilitates the focused analysis of parts of a large, complex network. We demonstrate the value of this approach by applying it to an investigation into the systems biology of telomere maintenance in the baker's yeast S. cerevisiae.

1.3 Telomere maintenance and BMH1/2

Telomeres are structures composed of the ends of eukaryotic chromosomes together with their capping nucleoprotein complexes. These structures appear to play a major role in the ageing process (Blasco, 2007; Cheung and Deng, 2008). Without the capping proteins, chromosome ends appear as double-stranded breaks to the cell's DNA damage detection mechanism, triggering a checkpoint response and ultimately cell cycle arrest (Longhese, 2008; Sandell and Zakian, 1993). With each cell division, telomeres shorten (Longhese, 2008). When telomere length falls under a certain threshold, the checkpoint response is triggered. This mechanism appears to contribute to the establishment of a cell's finite lifespan.

Saccharomyces cerevisiae is an excellent subject for studying telomere biology, because it is one of the simplest and most well-studied eukaryotic model organisms, and telomere biology is highly conserved across eukaryotes. In S. cerevisiae, an important component of the telomere maintenance mechanism is the protein Cdc13, which binds to the single-stranded DNA overhangs of telomeres. Cdc13 has two major functions: capping the telomeric DNA and recruiting telomerase (Garvik ; Nugent ), which is part of the telomere repair mechanism.

CDC13 is an essential gene, so it is difficult to characterize using knock-out mutants. However, there is a temperature-sensitive mutant called cdc13-1. This mutant has a wild-type phenotype below 26○C, but above this temperature telomeres become uncapped, the checkpoint response is induced and the cells stop dividing (Weinert and Hartwell, 1993).

Epistatic interactions between cdc13-1 and every non-essential gene of the S. cerevisiae genome have been studied (Addinall ) using a high-throughput synthetic genetic array (SGA) assay (Tong ). Addinall and co-workers identified a large number of genes which appear to genetically interact with cdc13-1. The biological function of many of these genes is already understood, but the role of others requires further investigation.

A particularly challenging problem is the phenotype of the two paralogues BMH1 and BMH2. The deletion of the gene BMH1 strongly suppresses the cdc13-1 phenotype, while the deletion of its paralogue BMH2 suppresses cdc13-1 rather weakly. BMH1 and BMH2 share 91.6% sequence identity. They both encode 14-3-3 proteins, a class of proteins which usually occur as dimers and which bind to phosphoproteins (Chaudhri, 2003). Members of the 14-3-3 family have a variety of different functions in eukaryotes, including directly modifying the functionality of their target proteins, mediating and controlling transport processes between the cytoplasm and different organelles and serving as scaffolds for interactions between different proteins (Tzivion ). An important role for BMH1 and BMH2 in the regulation of carbohydrate metabolism has also been suggested (Bruckmann ).

It is not immediately obvious why BMH1 and BMH2 behave differently despite their close homology. Clearly, the difference in phenotype between two such closely homologous genes cannot be understood by studying the genes in isolation; a systems biology approach is essential.

In order to investigate this problem, we used Ondex to integrate five publicly available data sources as well as a homology dataset generated from BLAST results. We then enriched this data with semantic links between the concepts from the different data sources. In addition, we developed a novel, view-based visualization approach which we used to analyse this large, complex dataset. This analysis produced several testable hypotheses.

2 METHODS

2.1 Data sources

We integrated six different data sources using Ondex. Genomics data, as well as the latest Gene Ontology (Ashburner ) annotations, were obtained from the Saccharomyces Genome Database (SGD) (Cherry ). A yeast regulatory network was acquired from Balaji ). A curated model of the yeast metabolic network was sourced from Herrgard ). A yeast protein–protein interaction (PPI) network and a network of known genetic interactions (GI) were taken from the BioGRID database (Stark ). The BioGRID data includes the GIs reported by Addinall ). Homology links between genes were created using BLAST (Altschul ) with an 85% sequence identity threshold (Table 1). Further information regarding the data sources can be found in the Supplementary Material.

Table 1.

Data sources used in this work

Data	Source	Version/date
Genome	SGD	11/02/2010
GO annotations	SGD	11/02/2010
Interactome	BioGRID	v2.0.61
Regulatory network	Balaji et al. (2006)	NA
Metabolic network	Herrgard et al. (2008)	v1.0
Homology	BLAST (id. >85%)	NA

NA = not applicable.

2.2 Integration

The source datasets were integrated to form a combined Ondex Saccharomyces knowledge network. A metadata model was designed for the knowledge network to capture the semantics of concept and relationship types found within the different datasets (Fig. 1). This model provides information about how certain concept types inherit features from one another (for example, an Enzyme is always a Protein), and allows such concepts to be treated as both concept types in the downstream workflow.

Fig. 1.

Simplified schematic of the data structure underlying the Ondex Saccharomyces knowledge network. Circles represent concept types, and lines represent relations between them.

Ondex parsers for each data source produce an Ondex-compatible representation of the data. For example, the BioGRID parser creates an appropriately typed concept for every reported gene, protein and RNA in the database, and an interaction concept of a corresponding type for every interaction. All genes, proteins and RNAs that take part in an interaction are linked to that interaction concept with a participates_in relation.

A mapping algorithm based on matching cross-references was used to find identical concepts in the Ondex graph. For example, the protein encoded by the ORF YDL220C from BioGRID matches a concept with the same accession number in the metabolic network. When the procedure identifies a group of matching concepts, it merges them into one single concept that captures all of the information that was represented by the group members. The procedure checks whether all concepts in a group have compatible types. If so, it automatically uses the most specific type present in the group for the merged concept. Otherwise it reports the inconsistency. For example, a concept cannot be both a Gene and a Pseudogene, but a concept can be both a Protein and an Enzyme, in which case Enzyme is the more specific type.

The semantic model reflects current understanding of molecular cell biology: concepts of type Gene have transcription relations linking them to mRNA concepts, which in turn connect to Polypeptide concepts via translation relations. Polypeptides can be part_of Proteins. Enzymes, a subtype of Proteins, may catalyze Reactions, which consume and produce Molecules such as Metabolites and Proteins.

Transcription_factors, a subtype of Protein, connect to Genes via regulation relations. Genes in turn are Nucleotide_features, which connect via adjacency or overlap relations, establishing genomic context. All Nucleotide_features can participate_in various Genetic_interactions. Similarly, Molecules, such as RNAs or Polypeptides can participate_in Physical_interactions.

Non-physical data can also be associated with existing concepts. Nucleotide_features can contribute to Biological_processes. Proteins can have a Molecular_function. Various parts of the knowledge network can also be associated with Publication concepts.

All of the parsers and mapping methods, as well as the metadata that were created for this work, are available as part of the Ondex suite, which is licensed under the GNU GPL v3, and is downloadable from http://www.ondex.org/. Further details regarding the integration process can be found in the Supplementary Material.

2.3 A novel visualization strategy

The graph produced by the integration is stored in the Ondex XML format OXL (Taubert ) and can be browsed in Ondex. The resulting network is large and complex, and contains many types of concepts and relations. In order to present a more succinct and biologically focused view of the data, we developed a plugin for the network visualization tool Cytoscape (Shannon ) called OndexView.

This new visualisation allows the user to define ‘views’ on the data. A view focusses on one type of concept. All concepts of that type found in the knowledge network are visualized as network nodes. The user can query the knowledge network for associations between the concepts. These associations will be visualized as edges between the corresponding nodes. Querying for associations is accomplished by specifying a set of metadata motifs. The instances of these motifs found in the underlying knowledge network are then used to create corresponding associations in the view. Metadata motifs are alternating sequences of concept types and relation types that begin and end on the same concept type (Fig. 2E).

Fig. 2.

Schema illustrating two examples for constructing associations based on metadata motifs. (A) The motif shown selects all paths from a polypeptide that is part of a transcription factor which regulates a gene that is transcribed to an mRNA which is translated into another polypeptide. (B) The motif from (A) is shown in context of the complete metadata structure. (C) A small example subnetwork to which the motif can be applied. Finding the motif (A) in the subnetwork, we identify three matching paths. Each element in these paths matches its corresponding element in the motif. (D) The view resulting from collapsing the paths identified in (C) according to motifs from (A) contains three edges; one for each matching path. (E) The motif selects all paths from a protein that modifies a process that produces a metabolite which is then consumed by another process back to another protein that modifies that process. Such a motif could be described as ‘precedence in a metabolic pathway’. (F) The motif from (E) is shown in context of the complete metadata structure. (G) A small example subnetwork to which the motif can be applied. Finding the motif (E) in the subnetwork, we identify two matching paths. Each element in these paths either matches or is a subtype of the corresponding element in the motif. (H) The views resulting from collapsing the paths identified in (G) according to motifs from (E) contains two edges; one for each matching path.

A modified depth-first search algorithm is used to extract motif instances from the underlying Ondex knowledge network. At each depth level, the algorithm checks whether the currently explored path matches the target motif. For example, a user interested in proteins and metabolic pathway relationships between them could specify a motif representing this type of interaction. Applying the algorithm using this motif produces a view on the graph, containing only those elements in which the user is interested (Fig. 2). It is possible to combine multiple motifs to form a view as long as the motifs share the same start point concept type.

2.4 Using OndexView with the Saccharomyces knowledge network for generating hypotheses

A semantically collapsed view of a complex network facilitates hypothesis generation by providing a simple representation of genes of interest and their interactions. A subnetwork consisting only of these genes and their neighbours can be generated, and inspected to identify edges of potential interest. Examination of the annotations attached to these edges and their adjacent nodes in the Ondex network provides links to available knowledge about the underlying biology. The semantically collapsed view also provides a way of perusing relevant literature in a focused manner, making hypothesis generation considerably more efficient than the alternative of trawling through large databases of publications.

Using OndexView, we defined five motifs over Gene concepts covering genetic and physical interactions, homology, regulation and metabolic precedence (Table 2).

Table 2.

Motif definitions used in this work

Association name	Motif
homologue	Geneis_homologueGene
metabolic path	GeneencodesPolypeptideis_part_ofProteinparticipates_actively_inProcessgivesMetabolitetaken_byProcesshas_active_participantProteinhas_partPolypeptideencoded_byGene
ph. interaction	GeneencodesPolypeptideparticipates_actively_inPhysical_Interactionhas_passive_participantPolypeptideencoded_byGene
regulation	GeneencodesPolypeptideis_part_ofTranscription_FactorregulatesGene
gen. interaction	Geneparticipates_actively_inGenetic_Interactionhas_passive_participantGene

We created a view using all five motifs, querying for the immediate neighbourhood of BMH1, BMH2 and CDC13. We then laid out the view, separating all neighbouring genes into groups: Exclusive neighbours of BMH1, exclusive neighbours of BMH2, joint neighbours of BMH1 and BMH2 and others.

3 RESULTS AND DISCUSSION

The work described here has four major outcomes: a knowledge network for the yeast S. cerevisiae; an algorithm for semantic collapsing of a network; a Cytoscape plugin for visualization of the knowledge network implementing this algorithm; and a set of hypotheses relating to telomere maintenance in S. cerevisiae, generated using this approach.

3.1 The Ondex Saccharomyces knowledge network

Using the Ondex data integration platform, we integrated five publicly available data sources covering various aspects of S.cerevisiae biology into a semantically enriched and interlinked knowledge network. The resulting knowledge network consists of a total of 240 964 concepts of 59 different types, connected by 754 391 relations of 28 types.

3.2 OndexView: a Cytoscape plugin

We implemented a new method for semantic simplification as the Cytoscape plugin OndexView, that uses views on the underlying data. Users can load integrated Ondex networks into OndexView and open views on them.

A user can invoke predefined views over the knowledge network or define custom views. A built-in motif editor is included in OndexView for this purpose. To apply a view, a concept type is selected from the knowledge network. The user then chooses one or more motifs for the selected concept type. OndexView extracts all the required data from the underlying Ondex graph and constructs associations according to the algorithm outlined in Section 2.3. The program then offers a query interface to the user, which can be used to focus on specific nodes (such as genes) or collections of nodes (such as pathways and complexes) and their neighbourhoods. Once the query has been processed, OndexView displays the selected view in the main Cytoscape window.

3.3 Telomere maintenance in S. cerevisiae

BMH1 and BMH2, despite their 91.6% sequence identity, have very different interactions with the temperature-sensitive telomere uncapping mutant cdc13-1. Our analysis of the Ondex Saccharomyces knowledge network, conducted using OndexView, led to the formulation of several hypotheses to explain the different interaction profiles of BMH1 and BMH2.

From the semantically collapsed graph, it is apparent that the neighbourhoods of BMH1 and BMH2 have only 23.5% of genes in common; BMH1 and BMH2 have completely separate sets of transcriptional regulators. This information is not readily apparent in the uncollapsed graph. Analysis of the node description fields reveals that 19.3% of the two genes' combined neighbourhood are cell cycle-related genes, while 10.1% are glucose metabolism related genes. Of the genes that physically interact with both BMH1 and BMH2, 14.3% are histone related (Table 3). An annotated screenshot of this view is available as Supplementary Figure S2.

Table 3.

Gene groups in the neighbourhood of BMH1 and BMH2 in the created view and their cardinalities

	Total	Cell cycle related No. (%)	Glucose metabolism related No. (%)	Histone related No. (%)
Neighb. of BMH1	68	15 (22.1)	7 (10.3)	4 (5.9)
Neighb. of BMH2	79	15 (18.9)	8 (10.1)	4 (5.1)
Intersection of BMH1/2 neighbourhoods	28	7 (25.0)	3 (10.7)	4 (14.3)
Union of BMH1/2 neighbourhoods	119	23 (19.3)	12 (10.1)	4 (3.3)

The over-representation of genes involved in regulation of the cell cycle in the joint neighbourhood of BMH1 and BMH2 suggests that BMH1 and BMH2 may function as cell cycle regulators. Notably, Rad53, a key element of the cell's checkpoint signalling pathway, interacts physically with Bmh2. Further, two genes that have previously been identified as suppressors of cdc13-1: BNR1 and CYK3, are also present in the neighbourhood of BMH1 and BMH2. The joint neighbourhood of BMH1 and BMH2 also contains a relatively large number of genes related to regulation of glucose metabolism, indicating that the pair of genes may also play a major role in the regulation of glucose metabolism (Bruckmann ).

3.4 Differential regulation of BMH1 and BMH2

The semantically collapsed view of the neighbourhood of BMH1 and BMH2 (Section 2.4) shows that Bmh2 physically interacts with the protein Rad53. After selecting Rad53 in the view, we can learn from its description field that Rad53 mediates the activation of the cell-cycle checkpoint (Schwartz ), thus interacting with the cdc13-1 phenotype. Examination of the interaction between Bmh2 and Rad53 reveals an annotation indicating that this interaction was originally reported by Usui and Petrini (2007). These authors showed that both Bmh1 and Bmh2 directly bind to the active (phosphorylated) Rad53 protein, thus enhancing its signalling effect. An edge between BMH1 and RAD53 is absent in the Ondex network, since Usui and Petrini experimentally verified only the physical interaction between Bmh2 and Rad53. However, they hypothesise that Bmh1 interacts with Rad53 in the same way that Bmh2 does.

Accepting Usui and Petrini's assumption, a significant difference in protein abundance during the time of checkpoint induction could explain the different effects of BMH1 and BMH2 deletions. The collapsed network neighbourhood of BMH1 and BMH2 shows clearly that the two genes are regulated completely independently from one another. Furthermore, according to their description fields in the network, two of BMH1's transcriptional regulators, ACE2 and SWI4, are known to be active during G1 phase (Andrews and Moore, 1992; McBride ). This observation leads us to conjecture that BMH1 could be more strongly expressed than BMH2 during the G1 phase, resulting in a higher abundance of Bmh1 protein during and after G1 phase.

We hypothesize that due to its independent transcriptional regulation, BMH1's products are present at higher levels than those of BMH2 at checkpoint time. Thus, deleting BMH1 would remove the majority of the Bmh1 and Bmh2 protein available as binding partners for Rad53, severely impairing the cell's checkpoint response in the face of uncapped telomeres. Removing BMH2, in contrast, removes a smaller proportion of binding partners for Rad53, thus resulting in a milder suppression effect on cdc13-1.

The differential regulation hypothesis has been summarized in Figure 3A, by generating a neighbourhood graph of BMH1 and BMH2 and removing all nodes except for those mentioned in the above discussion.

Fig. 3.

Screenshots from OndexView, summarizing the hypotheses presented. Edges represent physical interactions (blue), homology (red), regulation (green) and epistasis (yellow). Nodes represent genes, some of which are weak (light red) and some strong (dark red) suppressors of cdc13-1. Genes which show lethal or slow growing phenotypes upon deletion are marked with dashed outlines, as knowledge of their epistatic behaviour can be expected to be incomplete.

Experimental validation of the differential regulation hypothesis could be performed in several ways. Expression profiling of BMH1 and BMH2 over the course of the cell cycle could be performed. Consultation of pre-existing data has so far been inconclusive: a time course microarray performed by Spellman and colleagues showed BMH1 expression levels to be slightly higher than those of BMH2 throughout the the cell cycle and spiking 4-fold during G1 phase (Spellman ) (Supplementary Fig. S2). However, due to the low resolution of the assay, the observed peak is represented by only one datapoint, rendering its significance rather doubtful. Another possibility would be the examination of an ace2Δswi4Δ double mutant, which under the above hypothesis should replicate the bmh1Δ phenotype regarding suppression of cdc13-1.

3.5 Phosphorylation of Cdc13

A different hypothesis was formed after further examining the OndexView network neighbourhood of BMH1 and BMH2. The network shows a set of edges, indicating that Bmh1 and Bmh2 physically interact with Snf1, Snf4 and Gal83, which according to their node description fields, are members of the AMP-dependent kinase (AMPK) complex. Snf1 also interacts with Cdc13 and Stm1, another telomere-capping protein. The publication linked in the Ondex knowledge network reveals that this interaction is a phosphorylation, (Ptacek ).

Publications linked from the nodes reveal that the AMPK complex is a heterotrimer composed of the α-subunit Snf1 (the actual kinase), its activating γ-subunit Snf4 and a third component β-subunit that tethers Snf1 and Snf4 together. Sip1, Sip2 and Gal83 compete for the place of this third component (Jiang and Carlson, 1997). These three competing proteins have been shown to determine the AMPK's substrate specificity and its cellular localization (Lin ; Schmidt and McCartney, 2000). For example, the Gal83 variant of the AMPK complex has been shown to be able to enter the nucleus (Vincent ).

In the network, two edges exist indicating that Bmh1 and Bmh2 both physically interact with Snf1 (Elbing ). However, Bmh1 alone interacts with the γ-subunit Snf4 (Gavin ), while Bmh2 alone interacts with the β-subunit Gal83 (Krogan ).

The nature of these interactions is currently unknown, but 14-3-3 proteins have been reported to affect kinases in several different ways, including scaffolding and direct alteration of the target's function (Tzivion ). If Bmh1/2 are involved in scaffolding for the formation of different AMPK variants, then the deletion of BMH1 and the subsequent over-representation of Bmh2 dimers could favour the formation of Gal83-AMPK variants which can enter the nucleus to phosphorylate Cdc13. We, therefore, hypothesize that the phosphorylation of Cdc13 could potentially affect its temperature sensitivity and thus the cdc13-1 phenotype.

The phosphorylation hypothesis has been summarized in Figure 3B, by generating a neighbourhood graph of BMH1 and BMH2 and removing all nodes except for those mentioned in the above discussion.

Experimental validation of the phosphorylation hypothesis is more difficult. An examination of the phenotype of gal83Δ mutants could offer further insight. If such a test would corroborate the importance of the AMPK complex for the observed phenotype, one could examine whether in vitro phosphorylation of Cdc13 by the AMPK is possible. However, without further evidence we have to consider the effect of a potential phosphorylation of Cdc13 on its temperature sensitivity in particular to be mere speculation.

3.6 Impact of noisy data

Like any other data integration approach, the Ondex Saccharomyces knowledge network is limited by the state of knowledge contained in its data sources. There are two main types of errors that affect the system. (i) Incorrect information from the databases, such as curator errors, will also be present in the integrated network, unless detectable by Ondex's inconsistency checks as discussed in Section 2.2. For example, contradictions between data sources can be detected and rectified. (ii) Information that is missing from the source databases will also be missing in the integrated dataset. A particular problem in this respect is the lack of negative knowledge recording throughout the systems biology community. In many sources, it is not possible to decide if non-existence of a database entry is indicative of the subject being known not to exist or not being known to exist.

It is obvious that such problems also impact downstream analyses with OndexView. However, by enabling the user to review the evidence underlying the integrated data, she/he can be pointed at the original publications that can be consulted. For example, as described in Section 3.4, the publication linked from the physical interaction edge between Bmh2 and Rad53 clarified the circumstances around the non-existence of an edge between Bmh1 and Rad53. So it can be argued that performing visual analyses with OndexView on the Ondex Saccharomyces knowledge network also impacts on missing data, as it allows for clarifications and corrections.

3.7 Comparison to related works

While the Ondex Saccharomyces knowledge network in conjunction with OndexView's semantic simplification method shows parallels to previous data integration approaches, there are a number of important differences. Like the work presented in this article, semantic web approaches, such as YeastHub (Cheung ) store their integrated data in a machine-interpretable fashion, thus providing flexible platforms that can be adapted for various purposes. However, unlike OndexView, such systems require complex querying language constructs to access them. Data warehousing approaches such as SAMBA (Tanay ) and EcoCyc (Keseler ), on the other hand, are related to the presented work in that they collect the contents of heterogeneous data sources in one centralized repository. Unlike the Ondex Saccharomyces knowledge network, they offer web interfaces, which makes them very easy to query for standard use. However, they are based on less semantically rigorous data structures, rendering them less flexible. In summary, the Ondex Saccharomyces knowledge network combines aspects from both semantic web approaches and data warehousing approaches; featuring both their strengths, but also some of their weaknesses.

4 CONCLUSIONS

The data produced by high-throughput approaches has the potential to revolutionize our understanding of biology, but can do so only if the computational techniques necessary to visualize and analyse the data can scale with the amount of data generated. Data integration methods have proven to be a successful way to face this challenge. We have shown that a view-based approach can be a powerful tool to simplify the visualization of the complex knowledge networks generated by semantic data integration, without loss of the underlying information. The view-based approach provides concise visualizations tailored to providing only the information relevant to a particular investigation.

Biological phenomena such as the different phenotypes arising from deletion of BMH1 and BMH2 emerge from the interplay of several independent molecular biological networks. OndexView enables users to visualize not only these networks but also the ways in which they dovetail. This visualization serves as a starting point for the user, who can now easily explore genes, their various relationships and the underlying evidence, thus gathering inspirations facilitating the generation of testable hypotheses regarding the functions of these genes.

4.1 Outlook

There are a number of improvements that could be applied to OndexView and the presented knowledge network, as well as a number of ideas that build upon this work. The exploration of evidence trails behind the simplified edges in OndexView could be made more easily accessible. Rather than showing evidence in a textual representation in graph attributes, an option to visualize it graphically on demand would further increase the tool's usefulness. However, due to the limitations of the Cytoscape API, which renders graph views immutable, such new features will more likely be included in a re-implementation of OndexView for Ondex's own graph visualization frontend.

Furthermore, the Ondex Saccharomyces knowledge network could be enriched with probabilities on the connections between concepts, which could be inferred from the original data sources as well as the evidence coverage. Then, if more experience can be gathered on molecular pathways that potentially qualify for explaining observed epistasis effects, they could be generalized into a collection of semantic motifs. Instances of these motifs in the graph could be ranked according to their overall probability. The generation of such ranked lists could further assist biologists with the formation of hypotheses.