The challenge in bioinformatics is not integration by itself, which could be achieved with ad-hoc scripting, but to do so in a manner that is repeatable, customisable, and enables powerful queries and visualisations. Here we present a case study which should provide some valuable insights. We set out to integrate public datasets from Drosophila melanogaster, including pathway data from BioCyc and tissue expression pro les from FlyAtlas. Cytoscape and PathVisio were used for visualisation of networks and pathways. We used existing ontologies, such as BioPAX, where possible. We found situations where the existing standard had to be amended, for example we inferred new Xref classes using Identi ers.org IRI's. For the purposes of repeatability and provenance, we made every aspect of our system fully scripted. We discovered that Cytoscape visualisations work best to answer questions about the whole dataset, such as \Which submodules of the data are primarily active in one tissue?". The PathVisio tool is more adept at answering detail-oriented questions, such as \Which enzymes are gatekeepers to the glycolysis, and in which tissues are they active?". Our position is that semantic technologies provide more exibility; repeatability through scripting is of utmost importance; and multiple visualisation tools can be combined to address a broader range of biological questions. Availability: The integrated data is available as RDF dumps, cytoscape session and a SPARQL endpoint at http:// y.cloud.generalbioinformatics.com
Integration of diverse data is a recurring problem in life sciences. The challenge is not integration by itself, which could be achieved with ad-hoc scripting, but to do so in a manner that is repeatable, customisable, and enables powerful queries and visualisations. In short, data integration should be exible. We believe that semantic integration provides this exibility although it is not completely sufcient to deliver a \cradle to grave" solution. Here we describe a case study, which gives insights into the roads toward the desired exibility.
General Bioinformatics performs bespoke data integration and visualisation for customers, using customer data combined with public domain data and tools. Semantic web technologies allow integration and management of multifarious biological data. Biological questions are often di use in nature and require the ability to explore datasets. Thus, integration and knowledge management must be complemented with powerful visualisations. Semantic integration of data delivers the basis for generation of those visualisations.
We aimed to integrate and visualise public datasets concerning the fruit y
Drosophila melanogaster. We employed only open source software, namely
Virtuoso for the triple store, Apache Jena, Rd ib and R for data processing, and
Cytoscape [
The system is a collection of linked data, with the graphs stored in Virtuoso's triple store. We used the following datasets: pathways from BioCyc [6] and WikiPathways[8], tissue expression pro les from FlyAtlas, and gene and protein descriptions from FlyBase. In general, our approach is as follows: rst, data is converted into triple format if not already available as such and loaded into Virtuoso. Each datasource is imported into a separate graph to allow independent updating. The second step is to create a layer of inferences to \loosely" integrate the data. After loading the data, it is available for querying, though the data is not yet \linked". Linking depends on two aspects: the use of common predicates or ontologies, and the use of common IRI's. Both can be inferred using SPARQL CONSTRUCT queries. We relied on Identi ers.org to provide common resolvable IRI's for biological concepts.
We consider repeatability of primary importance, and therefore all data preparation steps are performed using scripts, either in Python using Rd ib, or Groovy using Jena.
What follows is a more detailed description of some of the integration steps and inferences we used.
FlyAtlas [9] is a collection of tissue speci c gene expression pro les of fruit ies. We obtained this data from open ydata.org in RDF form. (The RDF is shown in the Appendix as the black parts of Figure 2) For visualisation of FlyAtlas, both in PathVisio and in Cytoscape, it was necessary to calculate log foldchanges. Untransformed fold-changes map poorly to a linear range of colours since fold-changes between 0 and 1 correspond to the same dynamic range as fold-changes between 1 and in nity. We calculated log fold-changes using a separate script in the R statistical programming language and augmented the graph with these triples. The FlyAtlas data was combined with probe-to-gene mappings, using gene identi ers from FlyBase[7]. Similarly, linking to BioCyc, was achieved through creating triples using Identi ers.org IRI's for each ybase gene, using biopax:xref as predicate.
BioCyc is a collection of organism-speci c databases of gene annotations,
metabolic reactions, and pathways. For Drosophila this combined database is
called FlyCyc and is available as BioPAX [
WikiPathways [
The integrated data can be visualised di erently depending on the biological question to be addressed.
To get a global overview of the data, it is possible to load the whole network of enzymes and metabolites into Cytoscape, and project the expression pro les on top. Network visualisation in Cytoscape can give answers to the following type of question: which submodules of the metabolic network are tissue dependent? Cytoscape also allows one to apply di erent layout algorithms, and to see the connections between pathways.
The transfer from triple store to Cytoscape takes two steps. First, the backbone of the Cytoscape graph is generated using a SPARQL SELECT query with three output columns, representing source node, edge type and target node. The result is stored in Simple Interaction Format (SIF), which is imported into Cytoscape. In the second step, any number of node and edge attributes can be extracted with further SPARQL SELECT queries.
We used Groovy scripts to automate the procedure of querying our triple store and saving the data to SIF format. We then load this data into Cytoscape from within the same script, and generate network views, by using the Cytoscape RPC plug-in. (Figure 4 in the Appendix shows what the result looks like in Cytoscape and this Cytoscape session is also available for download from our website.)
Cytoscape provides a complete overview of a network, and it can be hard to get a detailed view of just a single pathway. PathVisio is better at displaying pathways, and better at answering pathway-oriented questions, such as: in which tissue is a given pathway active?
To answer this question, we also need to know which tissue the constituent genes are primarily active in. Simply assigning a gene to the tissue where it is most highly expressed is not statistically sound, since a gene could be just a fraction higher than average in one tissue due to random chance. Therefore we used an unsupervised clustering to group genes with similar expression patterns. We performed k-means clustering in R to assign genes to clusters. One characteristic of k-means clustering is that the number of clusters has to be chosen a priori. After several attempts we settled on 16 clusters, as that appeared the optimal number where clusters appeared to have a distinct, biologically interesting prole. Cluster numbers were added to flyatlas:CalculatedValues for the whole y (See also Figure 2, pink node in the Appendix below).
For example, cluster number 1 contains genes that are over expressed in the fatbody tissue. To see the e ect of this, we ranked pathways by enrichment of this cluster. Indeed, the resulting ranking shows a clear pattern, with many pathways related to fat metabolism (For further details please see Table1 in the Appendix). We created similar rankings for each of the 16 clusters, which are available for download from our website.
PathVisio allows for a more detailed view of what is happening in pathways. Examination of the glycolysis pathway image quickly leads us to nd tissue speci c-e ects occurring in CG10924, the rst step of the gluconeogenesis, which is highly overexpressed in fatbody. This is shown in the Appendix in Figure 1. Also SLC2A1, a glucose transporter with a gatekeeper function, appears to be regulated in a highly tissue-speci c manner.
During the course of this project we encountered some pitfalls, and it may be instructive to mention them here. For starters, we heavily relied on existing ontologies and actively encourage the use of existing standards such as BioPAX. BioPAX is a mature standard for pathway exchange. However, we found that we needed to add to the BioPAX standard. For visualisation in Cytoscape, we wished to create a bipartite graph representation of metabolic reactions, where each node corresponds either to an entity (metabolite or enzyme) or to a process (reaction or transport). Transformation of the BioPAX graph to a bipartite representation is made di cult by the fact that each connection between biochemical reactions and enzymes is mediated by a catalysis object. A catalysis object ties the enzyme (BP:controller) and the reaction (BP:controlled), plus a few more properties. Therefore, we inferred a third relationship directly between BioPAX BiochemicalReaction and enzyme named BP:controlledBy (See Figure 3, red parts). With this extra axiom it is now possible to extract the complete bipartite metabolic network in a single query.
We faced a generic problem with BioPAX EntityReferences.
EntityReference is a class with two datatype properties, one to refer to a bioinformatics database, e.g. \FlyBase", and one for a local identi er of that database, such as \FBgn0013682". The local identi er by itself is not enough to uniquely identify a biomolecule. Links across graphs based on EntityReference are complex operations that involve comparing two literals at the same time. Matching in SPARQL is more e cient when the nodes use an IRI. We inferred new bp:xref's for each existing bp:xref, using a standard IRI from Identi ers.org. (This is depicted in the Appendix in Figure 3, purple parts). Identi ers.org enables convenient identi er resolution, as well as direct URI-based linking. These are customisations for BioPAX that we nd useful, and we believe that wellwritten tools that consume BioPAX ought to be tolerant to this layer of inferences and adopt the fully exible open world assumption that is o ered by using semantic technologies. Although most BioPAX tools assume a closed world, we suggest that a more exible approach in these tools would be more appropriate and will encourage consumer use of BioPAX data.
In this study we had di culties using blank nodes in RDF. The problem we faced was to add additional calculated data to the yatlas:TissueVals nodes, which were all blank nodes. We were generating a graph with calculated values. We wanted to combine this with the existing graph, but it is close to impossible to merge two graphs on blank nodes - on import, new blank nodes get generated, leading to all kinds of problems in querying the data. The solution in our case was to construct new nodes with assigned IRI, and re-assert triples for each triple asserted for that blank node. The generated IRI is not resolvable, but it is identi able. And that makes it still preferable over blank nodes, which are neither resolvable nor identi able. We believe the use of blank nodes should be avoided for resources where new triples might be asserted in the future. Given that it is hard to predict the future use of data, our general advice is to avoid blank nodes altogether when possible.
Ideally, a fully exible system would achieve complete decoupling between data and tools, using only the semantic layer as intermediate. In essence, none of the output tools depend on any of the input tools: such that we could swap input data and each tool can be easily adapted to use fresh data. Thus, WikiPathways could replace BioCyc as a source of pathway data, and Ensembl could replace Flybase as a gene database. However, this is not always possible to achieve because of data dependencies of tools. For example, although PathVisio can read high-throughput data from our system, it can not yet read pathway data in triple form and relies on a direct link to WikiPathways. In this respect our system could be improved.
The exibility that we strive for is di cult to de ne, we believe that integration should be t for purpose and yet it is impossible to predict how data can be used and reused as technologies progress. Given this, integration should be as exible as current technologies allow. We feel strongly that semantic web technologies provide this exibility. In semantic reasoning, it is important to pragmatically focus on the problem at hand. We recommend to use simple identi er schemes such as Identi ers.org, and we readily expand the boundaries of BioPAX with new inferences when it serves our purpose. Semantic integration relies on common IRI's used by the subjects in the triples and the ontologies used in the predicates. In this project we used a loose integration approach and added layers of inferences on top of FlyAtlas and other datasets. The beauty of semantic technologies is that new assertions can be easily added in the RDF or OWL as needs dictate, giving a fully exible data representation.
In our opinion any analysis must be easily reproducible. Therefore, every part of the exercise should be scripted in a manner that allows not just for repeatability but allow tweaks. For example, If triple loading and construct queries were done manually, we would quickly lose track of the source of a given triple, and lose the ability to make small improvements to the process. In this respect, our choice of software was not arbitrary: we picked tools for their scriptability (and this requirement incidentally favours open source software). We created a custom API to enable scripted access to Virtuoso and scripted all Cytoscape visualisations using its RPC interface.
Finally, we want to emphasize the usefulness of combining visualisation tools with semantic technologies. In our experience biological questions require the transformation of data, re-purposing and re-spinning data into speci c contexts and delivering it in a format that can be visually understood. Data that has been integrated is by de nition more complex than the individual sources. The sum is greater than the parts. We found that data integration is not su cient when answering biological questions and that being able to visualise integrated data in the contexts of the biological questions is what makes the integration valuable. Furthermore, we believe that the use of multiple independent visualisation tools helps to address a broad range of biological questions. Too heavy reliance on a single piece of software would hurt this capability, on the other hand, being able to take the data and visualise for speci c purposes is a considerable asset. The Cytoscape network visualisation tool shines when used for whole-dataset questions. For example, it can quickly generate visualisations to answer questions such as \what fraction of enzymes have been measured in FlyAtlas?", \which submodules of the data are primarily active in one tissue?". PathVisio, on the other hand, is more adept at answering detail-oriented questions, such as \Which enzymes are gatekeepers to the glycolysis, and in which tissues are they active?".
For our clients we nd the semantic web approach to data integration is not only e ective, it is more e cient and more scalable. However, biological questions rarely come in the form that can be delivered in a single SPARQL query, and it is the visualisation of integrated data that has bene ted our clients in particular. 6. Ron Caspi et al. The metacyc database of metabolic pathways and enzymes and the biocyc collection of pathway/genome databases. Nucleic Acids Res, 38(Database 7. Susan Tweedie et al. Flybase: enhancing drosophila gene ontology annotations.
Nucleic Acids Res, 37(Database issue):D555{D559, Jan 2009. 8. Thomas Kelder et al. Wikipathways: building research communities on biological pathways. Nucleic Acids Res, 40(Database issue):D1301{D1307, Jan 2012. 9. Venkateswara R. Chintapalli et al.
Using yatlas to identify better drosophila melanogaster models of human disease. Nat Genet, 39(6):715{720, Jun 2007. 10. Martijn P. van Iersel et al. Presenting and exploring biological pathways with pathvisio. BMC Bioinformatics, 9:399, 2008.
Further details of the data, integration and visualisation are provided below. Our results are available on our public website. There is a SPARQL endpoint and downloadable RDF of the integrated data, a Cytoscape session containing the network visualisation, and a collection of pathway visualisations created with PathVisio. This and more details can be found at http:// y.cloud.generalbioinformatics.com
For visualisation of tissue enriched pathways we created pathway rankings for each of the 16 clusters. An example of the ranking for Cluster 1 is shown below: