Biologists and biotechnologists need to draw information from numerous distributed and heterogeneous resources, such as online biomedical databases, nomenclatures and specialised bioinformatics tools. These tasks can benefit significantly from semantic data federation with SADI Semantic Web services where multiple resources exposed through SADI services are accessed as a single virtual SPARQL-queriable database. We provide evidence in support of this premise by creating and testing a kit of public SADI services for a number of bioinformatics databases and programs, and by demonstrating how it can be used to serve real information needs of ecotoxicology researchers, by using the services to answer some model queries.
The future of semantic technologies depends on how quickly and
broadly they are adopted, which in turn depends on what value
these technologies deliver to end users. For semantic technology
researchers and engineers, this necessitates checking their ideas and
prototypes in real or realistic use scenarios driven by potential end
users. This paper presents the results of such “fieldwork”
testing the utility of a data federation approach based on Semantic
Web services for systems biology: we explore the use of SADI [
We would like to emphasise that this paper is not an introduction
to SADI. In particular, we neither discuss the technical details of
how SADI services can be discovered and invoked, although a brief
overview will be given in Section 1.2, nor compare SADI to other
Semantic Web services and data federation frameworks. For this,
we refer the readers to [
Many activities in biomedical research and biotech industry
require finding and combining information from multiple
heterogeneous and distributed resources. The scope of data requirements
for a biologist often includes many autonomous resources, such as
online biomedical databases, nomenclatures, ontologies, literature
and patent repositories, clinical databases and specialised analytical
Web services, such as biomolecular sequence search and alignment.
The state-of-the-art approaches to data integration –
datawarehousing and workflow scripting (see, e. g., [
We are advocating the emerging data federation paradigm where querying multiple heterogeneous distributed resources is as easy as querying a single database. The dynamic nature of research and biotech R&D activities implies that in many cases pre-programmed, e. g., form-based, querying is not enough and ad hoc querying is necessary. For ad hoc querying to be affordable in terms of labour it has to be self-service, so that non-programmer users, such as biologists or clinical research professionals, can formulate and execute queries without help from programmers. This may be possible if the querying is semantic, i. e., based on the use of shared formalised vocabularies and automatic application of knowledge written in the form of ontological axioms or logical rules. If querying is semantic, end users can formulate their queries in the terminology of their domain, without knowing how the underlying data is structured or specific mechanisms of access to the data. Our study contributes to the research dedicated to the realisation of this vision. 1.2
The SADI (Semantic Automatic Discovery and Integration)
framework [
First, SADI services can only consume input in the RDF format and
can only produce output in the RDF format. This completely
removes the problem of syntactic interoperability – any SADI service
can directly consume data produced by any other SADI service.
Second, every SADI service provides a special semantic description
that unambiguously defines what the service does, thus facilitating
the findability of the service by client programs when they need the
corresponding functionality. The description specifies what kind of
RDF graphs the service expects in the input and can process, in
terms of the concepts (properties and classes) that can be used in
the input RDF and, more importantly, specifies the concepts the
service can use to form the RDF graphs in the output.
The concepts used in descriptions of a set of SADI services,
together with related concepts from the underlying ontologies,
constitute a federated schema and can be navigated by users, including
non-technical users that understand the corresponding domain
terminology, to form meaningful queries over the network of available
services. Such a query can be expressed, for example, in SPARQL
and executed by a special query engine that will find SADI services
providing relations that may be useful to satisfy the query, identify
the data that can be sent as input to these services, and invoke the
services to retrieve more data, and so on, until it has enough data
to answer the query. Such engines can also apply ontological axioms
or rules to facilitate simpler and more intuitive queries. Currently,
there are two prototypes that implement this functionality:
opensource SHARE [
A typical scenario for publishing a resource as a set of SADI services is as follows: suppose, a user would like to access the Web site ClinicalTrials.gov – a registry of clinical trials – via SADI services, and wants to retrieve trials by disease names and extract information about trials, such as the names of the drugs studied. The publisher can do this with two services: getTrialsByDiseaseName and getDrugNameByClinicalTrial.
The first step of the resource publishing process is data modelling: it is necessary to decide how the data will be represented in RDF and, in particular, what relations and entities can be used. For illustration purposes, we assume that we use classes Name, Disease, ClinicalTrial and Drug, and properties diseaseIsTargetedByTrial, trialStudiesDrug, hasName. In our data model, a disease name can be linked to a trial, and a trial can be linked to the corresponding drug name as in the following RDF (in Notation3 syntax): :disease_name a :Name; :hasValue "Typhoid fever" . :disease a :Disease; :hasName :disease_name;
It is generally preferable to reuse some existing ontologies defining the required concepts, but when no such ontology is available, one can introduce the concepts in a small service-specific ontology and ensure that they have mnemonic names and/or descriptive textual labels, so that their meaning is apparent to users.
The next step is to define the input and output of the services in terms of the chosen entities and relations. The service getTrialsByDiseaseName will accept Disease objects with hasName attached to them, and in the output it will link the diseases to ClinicalTrial objects identified with their URLs in the ClinicalTrials.gov database via the property diseaseIsTargetedByTrial. The service getDrugNameByClinicalTrial will accept these ClinicalTrial URLs as input and link them via trialStudiesDrug to Drug objects linked to the corresponding names via hasName. In the schematic representation of the service descriptions below, the dashed and solid lines correspond to input and output specifications respectively. getTrialsByDiseaseName getDrugNameByClinicalTrial string
a
a
a hasName diseaseIsTargetedByTrial
a trialStudiesDrug hasName string The SADI framework uses OWL syntax to express the semantic descriptions. For example, the description of getTrialsByDiseaseName consists of the following two class definitions (in Prot´eg´e syntax): (Disease and (hasName some string)) as the service input class and (diseaseIsTargetedByTrial some ClinicalTrial) as the output class. 1.3
The discipline of ecotoxicology increasingly focuses on complex
interactions in biological systems which, on the technical side,
often requires an integrated analysis – a systems view – of Omics
data of different types, such as proteomics and genomics data (see,
e. g., [
There are several approaches to integrating biological data sets, and
some of them are based on Semantic Web. For example, projects
like Bio2RDF [
In the study presented here, we consider a number of research questions from the field of toxicology that require data and algorithm integration. To facilitate the integration of multiple resources which are required for obtaining insights on these research questions, we have created a kit of SADI services exposing these resources. In this paper we describe how the relevant data is modelled, how the corresponding SADI services work and how they can be used via SADI query engines to implement semantic federated querying of the resources. We provide three example SPARQL queries that demonstrate the potential utility of SADI-based semantic data federation for biological research.
Fish toxicology is a sub-field of ecotoxicology that studies the responses of fish to exposure to various pollutants, such as fertilisers and pesticides. Ultimately, such research is meant to facilitate the preservation of fish populations, develop more efficient aquaculture methods and also help to discover general biological mechanisms that may transfer to other species, especially humans. Typically, fish toxicologists want to know what sequences of molecular events are triggered in fish after they are exposed to chemicals of a certain type, and what organism functions and biological processes are affected. More specific examples of research questions might include “For all fish microarray studies, what complexes of biochemical reactions are commonly affected?”, “Does the exposure to chemical X affect the production of proteins with common structural properties?” or “What types of human diseases are known to be related to the genes affected by pesticide Y in fish brain?”. Answering questions of this kind often depends on the ability of researchers to analyse experiment results in the context of available biomedical knowledge. Here we describe the main types of databases and algorithmic resources that can be used in combination to get insights on fish toxicology-related questions.
Microarray experiment repositories. The toxicity of many
chemicals is manifested by affecting the expression of genes: some
genes are up-regulated (produce more RNA than usual), others are
down-regulated (produce less RNA). Consequently, an important
technique for learning about the effects of a chemical on an
organism is to measure the expression intensity for various genes by
conducting a DNA microarray experiment. Experiments of this kind
are also used for other tasks, such as measuring protein
quantities in biological samples. Large numbers of results of microarray
experiments are deposited in public online databases, such as
ArrayExpress [
Gene Ontology (GO) and model organism databases.
Meaningful interpretation of gene expression experiments often requires
mapping genes to GO annotations specifying molecular functions of
the corresponding proteins or biological processes they are involved
in. Experimentally derived annotations of this kind are available
from a number of genomic databases for well-studied (“model”)
organisms, such as human protein annotations at EBI [
This section describes how we model the data discussed in the previous section. Choosing appropriate ontological primitives and RDF modelling patterns is crucial for convenient and flexible querying as well as for the semantic interoperability of services.
Ontologies. In order to improve the re-usability of our SADI
services, wherever possible we reference existing upper and domain
ontologies. Table 1 lists the ontologies used by the SADI Web
services written for our experiment: The Semanticscience Integrated
Ontology (SIO) is an upper ontology providing a broad set of
general classes and properties, and is used extensively by many SADI
services. The Life Science Resource Name (LSRN) ontology
provides classes for records and identifiers from standard databases and
nomenclatures, such as lsrn:UniProt Record and lsrn:GO Identifier.
The Experimental Factor Ontology (EFO) [
Modelling patterns. The following figure shows a fragment of the federated schema for our SADI services, which can be used to design SPARQL queries. The predicates in brackets are the inverse of the properties represented by the connecting arrows.
sio:SIO_000028 sio:SIO_010000 has part organism lsrn:ArrayExpressExperiment_Record sio:SIO_000008 has attribute lsrn:ArrayExpressExperiment_Identifier sio:SIO_010015 Protein Sequence
sio:SIO_000231 efo:EFO_0000635 has input in organism part sio:SIO_000629 is subject of (is about) Our modelling facilitates many different Omics-related queries and is also future-proof because many SADI services, based on different but related data sources, can use the same or compatible modelling.
We have created almost 60 SADI services specifically for our fish toxicology use cases, that expose resources mentioned in Section 2. Where possible, we wrapped existing Web services implemented by database and tool providers, as SADI services to access live data. This ensures results are current and helps to avoid the maintenance cost associated with data mirrors. Descriptions of the services are provided at http://cbakerlab.unbsj.ca:8080/FISHTOX-SADIServices. We only briefly describe some of the services here.
ArrayExpress-based services. Several services which find experiments deposited in ArrayExpress and retrieve information about them, were created by wrapping RESTful services provided by ArrayExpress. The data was modelled using a combination of EFO (natively supported by the database) and SIO properties. The following figure shows the IO modelling for the service SpeciesName2AEExperimentalRecord that searches for ArrayExpress records by a species name: sio:SIO_010000 rdfs:subClassOf organism rdf:label
string microarrayso:has_instance efo:EFO_0000635 organism part efo:EFO_0002694 experimental process lsrn:ArrayExpressExperiment_Record lsrn:ArrayExpressExperiment_Identifier a a a a sio:SIO_000028 has part sio:SIO_000231 is input in The service implementation simply extracts the species name string from the input RDF, makes a request to the ArrayExpress search service and converts the resultant ArrayExpress record IDs to RDF.
vides a RESTful service for sequence search and alignment, so we created a SADI service that, given an amino acid sequence, finds Pfam records for protein domains similar to parts of the input sequence. The input class of the service is a ’protein sequence’ and the output class is defined in the hmmrso ontology as a class that ’has attribute’ some (lsrn:HMMR Alignment that (’is about’ some (’molecular site’ that (’is subject of’ some lsrn:Pfam Record)))) (here and throughout the rest of the paper single-quoted labels represent the corresponding identifiers from SIO). Similarly, SADI services for BLAST were created by wrapping NCBI Web services. Their inputs can be protein, DNA or RNA sequence strings, depending on the variant of BLAST. The output is defined using a special alignment class, in an approach similar to
HMMER3 services. For example, the output for the BLAST on proteins is defined in the blastso ontology as ’has attribute’ some (blastso:BLAST Alignment that (’refers to’ some (’protein sequence’ that (’is subject of’ some lsrn:NCBI NP Record)))).
Database ID mapping. Answering many queries requires
crossmapping of information from different databases, in one form or
another. For example, we may know a RefSeq ID of a protein
sequence, but retrieving Gene Ontology annotation requires UniProt
IDs of the protein. The RefSeq-to-UniProt mapping is already
available from RefSeq and, in general, biomedical databases often
reference other databases, so we only need an appropriate modelling in
order to implement the DB record mapping as a SADI service. Our
RefSeqNP2UniProt service takes an instance of (lsrn:NCBI NP Record
and (’has attribute’ some (lsrn:NCBI NP Identifier and (’has
value’ some string)))) as input and annotates it as an instance of
the following output class: (’is about’ some (’protein sequence’
and (’is subject of’ some lsrn:UniProt Record))).
Gene Ontology annotation. We created several SADI services
annotating proteins with GO identifiers of known molecular
functions they have and biological processes they participate in. The
data for several model organisms, such as humans, zebrafish and
house mouse, originating from the corresponding model organism
databases, are taken from the GO website [
In this section we present three example queries which address the types of questions pertinent to the analysis of fish toxicology experiments with DNA microarrays. However, the SADI services we have built, and the modelling we use, are not limited to these examples. Different combinations of our SADI services, together with multiple public SADI services developed for other purposes, can be used to answer other queries in contexts other than fish toxicology – the methodology we are using here is widely applicable to gene expression analysis in general.
We ran the queries with two SADI query engines – SHARE [
Experimentalists in fish toxicology need to compare their work with previous published experiments, which requires locating microarray experiments with related parameters. For example, a biologist may be interested in identifying genes whose expression has been measured in the tissue hypothalamus of the species largemouth bass in existing experiments. Currently, the main option is to use Web tools provided by experiment databases like ArrayExpress, which requires multiple searches and manual inspections of many experiments each of which may use different microarray platforms. SADI allows it to be done in a much easier way. A declarative SPARQL query whose simplified pseudo-SPARQL version is given below, expresses the question formally: SELECT ?experiment_id ?tissue_name ?platform_id ?gene_id FROM <http://cbakerlab.unbsj.ca:8080/fishtox/large-mouth-bass.owl> WHERE { ?org_class aeso:has_instance ?org_instance . ?org_instance a ncbi:NCBITaxon_27706 . # largemouth bass . ?org_instance ’has part’ ?org_part . ?org_part ’is input in’ ?exp_process . ?exp_process ’is subject of’ ?exp_record . ?exp_record ’has attribute’ [’has value’ ?experiment_id] . ?exp_record ’is about’ ?exp_process . ?exp_process ’has participant’ ?array . ?array ’is subject of’ ?array_platform_record . ?array_platform_record a lsrn:ArrayExpressPlatform_Record . ?array_platform_record ’has attribute’
[’has value’ ?platform_id] . ?array_platform_record ’is about’ ?array . ?array ’has part’ [rdfs:label ?gene_id] . ?org_part rdfs:label ?tissue_name .
FILTER regex(?tissue_name,"hypothalamus","i") .
}
Note that the RDF file large-mouth-bass.owl specified in the FROM clause contains the URI of the species to instantiate ?org class. The query was submitted to SHARE which resolved it by finding and calling the SADI service SpeciesName2AEExperimentalRecord to identify relevant ArrayExpress experiment records, the service AEExperiment2Platform to retrieve the corresponding microarray type (“platform”) records and the service AERecord2Microarray to extract microarray details, such as the DNA sequences. No understanding of the ArrayExpress semantic idiosyncrasies or data syntax was required to formulate the query. SHARE identified two microarray experiments satisfying the query.
Despite the fact that the SADI registry used in our experiments only contained services necessary for our experiments presented here, the query was quite hard for SHARE: it required an overnight run on a commodity quad-core server running both SHARE and the services, and practically all the time was spent inside the query engine itself. Hydra produced first answers in less than 4 minutes. When it was terminated after 1 hour, it had produced over 12,000 answers and was generating more answers. Less than 1/3 of the time was spent on behalf of Hydra itself, which is a good progress relative to SHARE. Moreover, a majority of the executed service calls were redundant, due to the current lack of corresponding optimisations in Hydra, which allows to estimate that Hydra’s performance will be several times better when such optimisations are introduced.
Gene Ontology annotation is very valuable for understanding toxicity of chemicals as it tells the toxicologist what (parts of) biochemical reactions are disrupted when a chemical affects the expression of a particular gene. However, such functional annotation is not directly available for many non-model species, such as largemouth bass, so biologists have to infer GO annotations based on sequence similarity with known genes in model organisms for which experimental evidence is recorded in public repositories.
The following query annotates ten most significantly affected genes
in largemouth bass treated with the pesticide dieldrin [
This query provides an interesting observation. The line ?refseq record ’is about’ ?protein sequence seems to duplicate ?sequence hit ’is subject of’ ?refseq record because ’is about’ is inverse to ’is subject of’ and the user could also postulate that the property ’is about’ is functional for RefSeq records. Unfortunately, we cannot replace the query with a more natural version by merging these two lines because neither SHARE nor Hydra currently support this level of expressivity.
In addition to tabular results, the RDF returned from the query captures detailed information regarding how candidate functions relate to the query inputs. For example: the quality scores for the alignments of similar proteins and the experimental evidence for the assigned functions of candidate orthologs.
This query also required an overnight run with SHARE. Hydra was able to execute it in less than 1.5 hour on the same hardware. 5.3
Another way to informatively characterise an affected gene – in addition to GO annotations – is to identify domains, i. e., subsequences with known biological functions, the corresponding protein contains. Often, this cannot be done for microarray sequences directly, because they are incomplete, and the use of ORF prediction tools is necessary to identify plausible protein sequences. Then, a sequence search and alignment procedure, such as HMMER, can be used to retrieve the domains. The following query implements this combination of tools and retrieves Pfam names for domains in proteins produced by the same ten largemouth bass genes as in the previous query: SELECT ?protein_domain_name FROM <http://cbakerlab.unbsj.ca:8080/fishtox/10genes.rdf> WHERE { ?DNA_chip_sequence ’is transcribed into’ ?RNA_sequence . ?RNA_sequence ’is translated into’ ?protein_sequence . ?protein_sequence ’has attribute’ ?alignment . ?alignment ’is about’ ?molecular_site . ?molecular_site ’is subject of’ ?pfam_record . ?pfam_record rdfs:label ?protein_domain_name } Three services are called by SHARE: the service DNA2RNATranscriber transcribing a DNA into an RNA, the service ORFPredictor predicting the ORF and translating the RNA to a protein sequence, and the service HMMR3PFamA retrieving Pfam domain IDs for subsequences of the protein. In several minutes the query engine returned the answer that a domain related to ribosomal activities was found on the gene UF Msa AF 100231, which accords with the results for Query II. The low coverage on genes (1/10) is not surprising given that our HMMER3 SADI service was configured to retrieve only strongly similar sequences. The service could be given more liberal parameters, but the service configuration functionality is not yet implemented in SHARE or Hydra. 6
We have demonstrated that semantic data federation with SPARQL querying of SADI services representing multiple bioinformatics databases and analytical tools can help to answer relevant research questions in fish toxicology. The SADI service kit we have created exposes the databases and programs in a semantically explicit way and enables diverse and powerful queries using query engines like SHARE or Hydra. When using this framework, an ecotoxicologist would not have to understand the semantics and technicalities of the underlying resources in order to formulate queries across databases and tools. He would not have to program any additional scripts of workflows as the query execution happens completely automatically.
Novel and emergent tools and algorithms that can be utilized for query resolution, such as new sequence alignment methods or reading frame prediction, can be incorporated into query resolution by implementing a SADI web service wrapper that uses the relevant service ontologies. The new services are registered to the SADI repository and these services will be incorporated without changing the query.
We conclude that altogether the approach promises to be practically relevant at least for Omics-based analyses of ecotoxicology data. We would like to emphasise that the queries we consider in this paper can be used in contexts other than ecotoxicology: e.g., the ability to search for relevant microarray experiments or Gene Ontology annotation of DNA sequences may be useful in many contexts where DNA microarray experiments are used. Currently, query composition required a semantic-web domain expert to compose SPARQL queries in close consultation with a biologist. Query composition required careful consideration of the semantic modelling, the available pool of services, and the performance limitations of the query-resolution engine. In many instances additional services were required to resolve new queries, so the design and testing of queries was an iterative process. We expect that query composition should become a faster and less painful process as the ecosystem of SADI tools and services matures. Client performance limitations are already being addressed with the development of the Hydra engine.
Regarding what could be improved in SADI and SADI query clients,
our current observations support most conclusions of [
We would like to discuss (i) in more detail here. In principle, nonprogrammer users can be trained to write SPARQL queries: for example, the third author, who is an undergraduate Physics student, actively participated in composing and debugging the queries. A query is typically composed incrementally: the user write a few lines initially, runs a query engine, assesses the results, adds a few more lines, an so on. This process will be significantly simplified when tools for extraction of semantic schemas from service descriptions and ontologies are implemented: such tools facilitate easy look-up of classes and properties that are meaningful in a particular query context.
However, even simplified by schema navigation, SPARQL querying
is not sufficient to fully realise the vision of self-service querying
because it is difficult to expect that it can appeal to the mass
user. More friendly graphical or keyword-based query interfaces, as
in [
The SADI framework has been compared to related approaches
in a number of publications – see, e.g., [
In [
We would also like to mention a case study [