The Linked Data paradigm involves practices to publish, share, and connect data on the Web. Thus, it is a compelling approach for the dissemination and re-use of scienti c data, realizing the vision of the so-called Linked Science. However, by just converting legacy scienti c data as Linked Data, we do not fully meet the requirements of data re-use. Scienti c data is evolving data. To ensure re-use and allow exploitation and validation of scienti c results, several challenges related to scienti c data dynamics should be tackled. In this paper, we deal with the publication of diachronic life science linked data. We propose a change model based on RDF to capture versioned entities. Based on this model we convert legacy data from biological databases as diachronic linked data. Our linked data server can assist biologists to explore biological entities and their evolution by either using SPARQL queries or navigating among entity versions. All services are publicly available at http://diana.imis.athena-innovation.gr/lod.
As technology advances in scienti c hardware (e.g., sequencers), easing the
creation and consumption of large volumes of scienti c data, and as more scienti c
datasets are published and become available for potential data users, the
current trend moves towards open science data. Linked Data is a powerful and
compelling approach for spreading and consuming scienti c data, involves
practices how to publish, share, and connect data on the Web, and o ers a new way
of data integration and interoperability. The driving force to implement Linked
Data spaces is the RDF technology1. The basic principles of the Linked Data
paradigm is (a) use the RDF data model to publish structured data on the Web,
and (b) use RDF links to interlink data from di erent data sources. Linked Data
technologies have given rise to the Web of Data: \a Web of things in the world,
described by data on the Web" [
However, by just converting legacy scienti c data to Linked Data (LD), we do not fully meet the requirements of data re-use. To ensure the e ciency in data 1 http://www.w3.org/RDF/ re-use and allow both the exploitation and validation of scienti c results, several challenges related to the dynamics of scienti c data must be tackled. Scienti c data is diachronic data. Thus, users (a) should have access not only to up-to-date scienti c LD, but to any of the previous versions and (b) should also be able to track the changes among versions, as well as their causes and e ects. In this paper, we describe our work on publishing diachronic life science data, and more speci cally, genomic, experimental and bibliographic data related to miRNA molecules. We have integrated data from well-known databases, tracked their evolution, and published them as diachronic linked data. We propose a change model based on RDF to capture versioned entities. Our linked data server can assist biologists to explore biological entities and their evolution. 2
MicroRNAs The discovery of microRNAs (miRNAs) in early 2000s has dramatically changed the view of biologists about the role of the so-called junk DNA. miRNAs bind themselves to message RNA (mRNA) transcripts, called targets, and regulate their expression. Knowledge of miRNA targets is very important for therapeutic uses. For example, such knowledge could be used to down-regulate genes by introducing arti cial miRNAs into the cells. Since the terms \miRNA" is nowadays used in a wide scope, it is common to distinguish between hairpin miRNAs and mature miRNAs, or just hairpins and matures from now on. The former signi es a precursor of the latter. A hairpin can actually be processed into several matures. Matures bind themselves to transcripts and prevent the creation of functional ribosomes.
Diachronic miRNA data During the past years, we have developed DIANA tools2, a set of advanced Web applications supporting a series of sophisticated work- ows enabling users with no strong background in informatics to perform advanced multi-step functional miRNA analysis. Under the hood of DIANA tools lies a relational database storing information for key entities of the miRNA domain, such as hairpins, matures, transcripts, genes, KEGG pathways, and publications. Information has been aggregated from a variety of sources, i.e information concerning the miRNAs comes from the miRBase database (up to v.18)3, gene related information from the Ensembl database (v.69)4 and molecular interaction and reaction networks from KEGG pathways database5.
The miRBase database is a searchable database of published miRNAs and maintains information for 18; 589 hairpins and 21; 881 matures. It also maintains a list of les that record successive versions along with the changes between them. By examining all these les, a number of changes are observed. Analytically, we observe the following changes: (a) NAME/SEQUENCE : change 2 http://diana.imis.athena-innovation.gr/DianaTools/index.php 3 http://www.mirbase.org/ 4 http://www.ensembl.org/index.html 5 http://www.genome.jp/kegg/pathway.html over the values of name and sequence respectively for both hairpins and matures, (b) NEW/DELETE : creation and deletion of a hairpin/mature entry, (c) FORWARD : merging of multiple hairpin entries into a single one, (d) ADD HAIRPIN PARENT/REMOVE HAIRPIN PARENT : creation/deletion of the relationship among a pair of hairpin and mature. Being able to track the changes of the miRNA and having access to older versions it is a crucial information when working with miRNAs. For example, the hairpin MI0001364 was introduced into miRBase at v5.0 with name dre-mir-10b and at v7.0 its name changed to dre-mir10b-1. Having access to all possible names, allows us to better track bibliographic references independent of its latest name. 3
A challenging problem in LD publishing is how to deal with evolving data. In the general case, changes are complex objects that interact with data objects, they can have structural, temporal and probabilistic characteristics, thus they cannot be handled as plain di erences neither transformation operations between datasets. This is evident especially in life science LD, since by just converting legacy scienti c data as LD, we do not fully meet the requirements of data reuse. To support the above requirements, we adopt versioned RDF entities and version properties, and we also model changes as rst class citizens, i.e. RDF resources, in the LD itself. Based on our approach, both query and navigation services can be developed on top of diachronic LD sets. 3.1
Versioned entities and properties To represent up-to-date RDF descriptions, we use a general URI that is based on the class of the resource and an id that uniquely identi es the resource: http://.../resource/class-name/resource-id. By using such URIs, one can retrieve the RDF description for the current version of a requested RDF resource. To retrieve the RDF description for an RDF resource in a previous version of the LD set, we use version timestamps to append the general URIs: http://.../resource/class-name/resource-id/timestamp. For example, the URI http://.../resource/matures/MIMAT0009477 retrieves the current version of the mature with id MIMAT0009477, while http://.../resource/matures/ MIMAT0009477/17 the description of the same mature for v17.
While versioned URIs support the navigation among versioned entities by following those URIs, we use four version properties to facilitate the querying of diachronic LD set: (a) :label, with the constant value "now", (b) :version, with version timestamp values, (c) :previousVersion, and (d) :nextVersion. The rst property can be used in SPARQL to provide access only to the current version of the RDF resources, while the second can be used to form queries that span several versions of the LD set. The other two properties are used to ease the navigation among di erent resource's versions back and forward, respectively. A key requirement for diachronic LD is the need to trace the origins of data and transformations that occurred. Capturing the changes themselves and having querying capabilities on them, actually give us the possibility to step back and examine how and when the changes took place. In other words, it opens the key challenging problem of preservation and provenance support. Our approach uses the RDF schema itself to represent changes. We propose two di erent strategies depending whether the change occurs in an RDF triple that involves a property with a literal object or a property connecting two resources.
Property to Literal In this case, the literal value is substituted by a new value. We model this change by adding to the RDF schema an extra property :change, which is a property to literal and describes the type of change. Figure 2 shows an RDF subgraph of the miRNA LOD, where a change occurs to the property diana:name of the hairpin MI0000069 at v2.0. On the left side, we see its RDF resource at the previous miRBase v1.5, which also contains the old name value hsa-mir-15a (depicted in red), while on the right side we see the RDF resource of the hairpin at miRBase version 2.0 with the new name value hsa-mir-15a and the extra property diana:change with literal value the type of change {in this case "NAME"{ that models the change itself (depicted in green). Property to Resource In this case, the change is an RDF resource itself that links to the two resources of the property (the subject and the object) that changed. This \change" resource apart from links to the corresponding resources, it contains also a time property, showing at which time point the change occurred. Other characteristics, like probabilistic ones, can easily be modeled by adding extra properties to the \change" resource. Figure 3 shows an RDF subgraph of the miRNA LOD when a change occurs at property diana:producesMature between the hairpin MI0000769 and the mature MIMAT0009199 at v17. At the top of the Figure 3 we see the two connected RDF resources at the previous v16 (depicted in red). At the bottom, we see the two resources at v17 where the property diana:producesMature has been dropped. To proper model proper this change, we extend the RDF graph with an instance of the class RemoveHairpinParent (depicted in green) that contains links to the respective resources at v17.
All the curated information related to miRNAs for DIANA tools is stored in a
relational database, but a Linked Data representation is necessary and useful for
biologists. Such a LD representation will be open and available to everyone for
reuse, exploration and validation. Moreover, it can be used for inference and can be
potentially be part of an automated interpretation tool of the role of non-coding
RNAs. Due to the internal relational storage of the data, we adopt the \virtual
RDF" approach: accessing a non-RDF database using an RDF view. Such an
approach enables the access of non-RDF, legacy databases without having to
replicate the whole database into RDF. We use the D2R server [
In order to show the abilities of our proposed model and its exibility, we present a number of SPARQL queries. The queries are categorized into the following types: (a) retrieval of up-to-date entities, (b) retrieval of diachronic entities and (c) change exploration. The retrieval of up-to-date entities requires the triple pattern ?e diana:label "now" as mentioned at Section 3.1. The following SPARQL query is an example of this type, retrieving the targets of miRNAs predicted by microT-ANN and showing the whole biogenesis path. Thus, we get the hairpin's URI, the URI of its produced mature and the transcript's URI that is targeted by the mature.
SELECT ?hairpin ?mature ?target WHERE { ?hairpin rdf:type diana:Hairpin. ?hairpin diana:label "now". ?hairpin diana:producesMature ?mature. ?interaction rdf:type diana:Interaction. ?interaction diana:hasMature ?mature. ?interaction diana:hasTarget ?target.
?interaction diana:application "microT-ANN (v4.0)". }
To retrieve diachronic data we need just to specify the period of interest. For example, if we want to reproduce the previous query, but for a speci ed previous miRBase version, e.g version 13.0, we have the following SPARQL query : SELECT ?hairpin ?mature ?target WHERE { ?hairpin rdf:type diana:Hairpin. ?hairpin diana:version "13.0". ?hairpin diana:producesMature ?mature. ?interaction rdf:type diana:Interaction. ?interaction diana:hasMature ?mature. ?interaction diana:hasTarget ?target. ?interaction diana:application "microT-ANN (v4.0)".
Finally we present two examples of SPARQL queries that show the evolution of data. The rst one retrieves the hairpins that changed their name throughout their lifespan and returns the hairpin's URI, the miRBase version the name changed, the old name value and the new name value: SELECT ?hairpin ?version ?old_name ?new_name WHERE { ?hairpin rdf:type diana:Hairpin. ?hairpin diana:change "NAME". ?hairpin diana:version ?version. ?hairpin diana:name ?new_name. ?hairpin diana:previousVersion ?hairpin_prev. ?hairpin_prev diana:name ?old_name. In order to retrieve the old name values of the Hairpins, we can access the previous version of the hairpin via the property diana:previousVersion. The second one retrieves all currently not-related hairpin and mature pairs that used to be related in the past and returns from left to right the hairpin's URI, the miRBase version the property diana:producesMature dropped and the mature's URI : SELECT DISTINCT ?hairpin ?version ?mature WHERE { ?connection rdf:type diana:removeHairpinParent. ?connection diana:version ?version.
FILTER (?version <= "18"). ?connection diana:hasHairpin ?hairpin.
?connection diana:hasMature ?mature.
Using the property diana:version in FILTER we can retrieve all changes in the relationship between hairpin and mature miRNAs before that version. Note also that since the property diana:producesMature connects two resources, a change in that relationship-type property we model it as a new RDF class. 5
This work is an extension of [
In this paper we presented our work on the publication of diachronic life science
LD. More speci cally, we aggregated data from well known databases related to
miRNA molecules and exported as LD. The curated data contains information
from experimental observations, as well as change and version information about
the miRNA entities. The miRNA LD server can assist biologists to explore
biological entities, navigate between versions of the resources and via the SPARQL
endpoint it provides access to applications for querying historical miRNA data
and tracking their changes. As future work, we plan to interlink our linked data
set with the linked data provided by the EBI RDF Platform [