PIBAS FedSPARQL: a web-based platform for integration and exploration of bioinformatics datasets

Abstract

Background: There are a huge variety of data sources relevant to chemical, biological and pharmacological research, but these data sources are highly siloed and cannot be queried together in a straightforward way. Semantic technologies offer the ability to create links and mappings across datasets and manage them as a single, linked network so that searching can be carried out across datasets, independently of the source. We have developed an application called PIBAS FedSPARQL that uses semantic technologies to allow researchers to carry out such searching across a vast array of data sources.

Results: PIBAS FedSPARQL is a web-based query builder and result set visualizer of bioinformatics data. As an advanced feature, our system can detect similar data items identified by different Uniform Resource Identifiers (URIs), using a text-mining algorithm based on the processing of named entities to be used in Vector Space Model and Cosine Similarity Measures. According to our knowledge, PIBAS FedSPARQL was unique among the systems that we found in that it allows detecting of similar data items. As a query builder, our system allows researchers to intuitively construct and run Federated SPARQL queries across multiple data sources, including global initiatives, such as Bio2RDF, Chem2Bio2RDF, EMBL-EBI, and one local initiative called CPCTAS, as well as additional user-specified data source. From the input topic, subtopic, template and keyword, a corresponding initial Federated SPARQL query is created and executed. Based on the data obtained, end users have the ability to choose the most appropriate data sources in their area of interest and exploit their Resource Description Framework (RDF) structure, which allows users to select certain properties of data to enhance query results.

Conclusions: The developed system is flexible and allows intuitive creation and execution of queries for an extensive range of bioinformatics topics. Also, the novel "similar data items detection" algorithm can be particularly useful for suggesting new data sources and cost optimization for new experiments. PIBAS FedSPARQL can be expanded with new topics, subtopics and templates on demand, rendering information retrieval more robust.

Keywords: Bioinformatics; Data integration; Data mining and information retrieval; Federated SPARQL query; Ontologies.

Fig. 1

PIBAS FedSPARQL architecture overview. The architecture consists of two main layers: query engine and user interface. The user interface enables users to construct simple and advanced queries and view the results of their execution. The query engine preforms a series of demanding processes that needs to be done before queries can be executed. The main query engine component, Data Source Manager, scans the local DataSources ontology, reads the user’s input and passes the information through the Query preparation component to the SPARQL query runner component, where the queries are executed. The Dataset projection component plays a role in the “Dynamic query filter” feature, allowing users to easily discover the structure of underlying datasets included in Federated SPARQL queries. The Detecting Similar Data Items component identifies similar data items from results retrieved after running predefined queries or queries extended with new datasets

Fig. 2

Representations of basic relations in the DataSources ontology in the Protégé editor a) Topic Biology b) Subtopic BiologyTarget c) Template “Found targets for the drug” and some of its properties d) PIBAS/CPCTAS dataset instance. This figure shows screenshots of the local ontology DataSources in the Protégé ontology editor [50]. The ontology contains information about initiatives and datasets included in predefined Federated SPARQL queries. Each dataset in the ontology is represented as an instance of a certain class. The object property conectedWith connects dataset instances with template instances. Every Subtopic class instance is connected with a Template class instance through the object property hasTemplate. Every Topic class instance is connected with a Subtopic class instance through the object property hasSubTopic

Fig. 3

Predefined query of Template2 for its pre-selected datasets. This figure shows the predefined Federated SPARQL query of the template “Find targets for the drug”. This query covers the PIBAS/CPTAS, Drugbank/Bio2RDF, ChEMBL/EMBL-EBI and BindingDB/Chem2Bio2RDF datasets. All predefined Federated SPARQL queries in the local DataSources ontology contain “%s” characters which represent objects values that will be replaced with the keyword entered by the user. The last “%s” character will be replaced with a particular pattern query if a new dataset is added using the “Add new dataset” feature

Fig. 4

Running of predefined query in PIBAS FedSPARQL a) Initial user interface b) Results after running predefined query. The initial user interface allows users to create queries in a very simple way by selecting a (sub)topic, template and entering a keyword. By clicking on the “Run query” button, the predefined Federated SPARQL query is executed and users receive results in the form of a table. The first column shows the retrieved results as URI or string. The second column displays the data source and initiative name. The icon in the top-right corner of the table shows statistical information about the retrieved data, including data source name, initiative name and the number of obtained data items per data source

Fig. 5

Adding new dataset to predefined query. This figure shows the pop-up window that allows users to incorporate any new dataset not included in the predefined list of datasets for an existing template. Users need to enter the dataset name, initiative name, dataset link, a comment, the endpoint URL, pattern query and the dataset properties most relevant for the selected template and topic. Finally, they need to click the “Add” button to complete the action. Conversance with SPARQL and the underlying ontology is necessary for this step

Fig. 6

Rewritten predefined query after adding new dataset. This figure shows the rewritten predefined Federated SPARQL query of the template “Find targets for the drug” after incorporating a new test dataset

Fig. 7

Result set after adding new dataset to predefined query. This figure shows the results in a table after executing the rewritten predefined Federated SPARQL query. The results table has the same layout as in Fig. 4

Fig. 8

Accordion elements for dynamic query filtering a) List of predicates for PIBAS/CPCTAS dataset b) List of predicates for BindingDB/Chem2Bio2RDF dataset. This figure shows the dynamic accordion elements for the PIBAS/CPCTAS and BindingDB/Chem2Bio2RDF datasets. The accordion elements contain a list of dataset properties which are dynamically created according to the template “Find targets for the drug”. Each property listed in an accordion element is hyperlinked to a web page with its description. The same applies to all datasets used in Federated SPARQL query. Users can select their desired properties and add them to the query by clicking on the “Add to query” button

Fig. 9

Generated star-shaped query for BindingDB/Chem2Bio2RDF dataset after dynamic query filtering. This figure shows the star-shaped SPARQL query created for the Binding/Chem2Bio2RDF dataset after adding the properties http://chem2bio2rdf.org/bindingdb/resource/CID_GENE and http://chem2bio2rdf.org/bindingdb/resource/uniprot to the query

Fig. 10

A sample result table after dynamic query filtering. This figure shows the results of dynamic query filtering. The results are organized by source (PIBAS/CPCTAS and BindingDB/Chem2Bio2RDF) and displayed in a paginated table. They can be sorted and filtered in order to refine the query result and show only the most relevant information

Fig. 11

Process of string transformation. The process of string transformation implies conversion and filtering of a string. Initially, the string is converted to lower case. Then it passes through regular expression filtering to extract alphabetic and numeric characters [a-z, 0–9]. The string is then purified by eliminating words that are in the list of stop words. This list contains high-frequency words with relatively low information content (function words and pronouns). Finally, suffix removal is performed by applying Porter‘s Stemming Algorithm [52]

Fig. 12

Similar data items (URIs) obtained by our algorithm after adding a new dataset. This figure shows similar targets detected in the results retrieved after adding a new dataset to the “Find targets for the drug” template and running the rewritten predefined Federated SPARQL query. The results are shown in the form of a table on a new web page

Fig. 13

Matching results using methods Predicates selected and Predicates not selected. This figure shows a graphical representation of data from Table 6. The graphic contains the number of similar data items obtained using two approaches, Predicates selected and Predicates not selected, and similarity matching result based on human judgment (0 means that no matching exists, 1 means that a matching exists)

Fig. 14

Results of usability evaluation obtained from our questionnaire. This figure shows the final outcome of the survey carried out in cooperation with RC staff. For this survey, the six-item Likert scale-based System Usability (SUS) questionnaire was used. In order to numerically analyze the survey results, the Likert scale responses were translated to numbers using the following five-point scale: 1 = strongly disagree; 2 = disagree, 3 = neutral; 4 = agree; 5 = strongly agree. Based on the questionnaire outcome, average values (AVG) and standard deviation values (STD) were calculated and graphically presented