This paper presents the rules used within the Open PHACTS (http://www.openphacts.org) Identity Management Service to compute co-reference chains across multiple datasets. The web of (linked) data has encouraged a proliferation of identi ers for the concepts captured in datasets; with each dataset using their own identi er. A key data integration challenge is linking the co-referent identi ers, i.e. identifying and linking the equivalent concept in every dataset. Exacerbating this challenge, the datasets model the data di erently, so when is one representation truly the same as another? Finally, di erent users have their own task and domain speci c notions of equivalence that are driven by their operational knowledge. Consumers of the data need to be able to choose the notion of operational equivalence to be applied for the context of their application. We highlight the challenges of automatically computing co-reference and the need for capturing the context of the equivalence. This context is then used to control the co-reference computation. Ultimately, the context will enable data consumers to decide which co-references to include in their applications.
Within the life sciences there has been a proliferation of databases published,
with the 2013 NAR database issue listing over 1,500 [
While there have been attempts to provide global identi ers for a concept,
e.g. with life sciences identi ers [
Then there is the problem of what does it mean to be the same; just
representing the truth is not a tenable position. As demonstrated by Halpin et al [
there are many interpretations of the owl:sameAs relationships that exist in the data. Life sciences datasets contain complimentary and overlapping data, modelled with di erent levels of granularity depending upon the purpose of the data capture. Therefore, when we refer to the data being about the \same" concept, we are not looking for identical representations but rather saying that these two complimentary records can be treated as being operationally equivalent. However, the notion of operational equivalence depends upon the use to which the data will be put and thus can only be decided by the user or the application they use.
Linked data allows a user to navigate their way through the web of data by
following links from one resource to a related resource. Consequently not every
pair of datasets are linked since you can navigate your way through the web
of data. However there are many scenarios where you need to know all of the
equivalent URIs. For example, to power a linked data integration platform such
as the drug discovery platform being developed by the Open PHACTS project
[
In order for scientists to trust mappings, they need to understand the context of the equivalence claim and who is making the claim. By providing the context { in scienti c terms { together with the provenance of the mapping { how it was made and by whom { the scientist can understand the notion of equivalence captured and make an informed decision about whether to include it in their application.
In this paper we { Identify the challenges of identity co-reference across datasets (Section 2); { Discuss the metadata required to describe a dataset and capture the context of its links to other datasets (Section 3); { Present the rules used to control co-reference computation and their usage in the Open PHACTS Identity Management Service (IMS) (Section 4). 2
Information relevant for drug discovery research is sourced from a variety of
overlapping datasets. For example, information about drugs can be retrieved
from DrugBank [
modelled with a di erent focus, and have their own identi er scheme, when can we say that two records are truely equivalent? In some cases it is straightforward. An entry in ChemSpider and an entry in ChEMBL that share the same InChI will report about the same chemical, e.g. \imatinib mesylate". However, when we consider the drug entry in DrugBank, e.g. \Gleevec", there can be multiple InChI entries associated, e.g. the \gleevec" entry contains the InChI for both \imatinib" and \imatinib mesylate". In this case, are the records the \same". For a scientist interested in \gleevec" they would be, but for someone interested only in \imatinib mesylate" perhaps not.
Many datasets contain links to other related datasets. For example, UniProt
includes links to several related datasets. However the nature of these links are
not captured; in the case of the RDF export of UniProt they are all stated as
rdfs:seeAlso. It is therefore hard to automatically reuse such links due to the
di ering natures of the datasets and meaning of the link. A case in point would
be the relationships stated between UniProt { a protein sequence dataset { and
Protein Data Bank (PDB) [
For users and applications to trust and reuse equivalence relationships, they need to understand what notion of equivalence is being claimed. UniProt, in their RDF export, weaken their links to rdfs:seeAlso to avoid making inaccurate claims, but this reduces the knowledge conveyed. At the other extreme, the datasets in the Linked Data Cloud tend to be very relaxed about their claims of \equivalence" and widely use, or misuse, the predicate owl:sameAs; typically they do not intend the strict semantics of owl:sameAs. As such, these links need to be used with caution. Such context will enable applications to choose which links to include. For example, for the vast majority of drug discovery research { almost all { it would be acceptable to use \equivalence" links between gene entries in Entrez Gene and protein entries in UniProt as genes are often used as proxies for the protein that they encode. For example, when searching for data about a target, it is common to enter the gene name as the search term. However, there are those who would require that such links are not included, e.g. working in very niche specialisms or on edge cases. 3
For e ective linking between datasets, and to enable trust in their use in applications, it is essential to understand what has been linked and how. This requires descriptions of the datasets and the links themselves.
A dataset description is essential for data discovery and to enable consumers to use the data. It is a means to provide core metadata about a dataset, e.g. its title, description, and license information. It can also convey information about how to access the dataset, e.g. a SPARQL endpoint location, how the data is modelled, i.e. which vocabularies have been used, and key statistics about the data, e.g. number of triples, number of subjects, etc.
The Vocabulary of Interlinked Datasets (VoID) [
However, VoID does not prescribe which properties must be provided and
those that are more optional. This makes the general re-use of VoID dataset
descriptions di cult as there is no guarantee that the information you need will
be provided. For example, for the pharmaceutical companies involved with Open
PHACTS it is essential to understand the licensing restrictions of the dataset,
but this information may not be present in the dataset description. There is also
no notion of capturing the version of a dataset, which is essential to know when
linking between datasets. To overcome these challenges in the Open PHACTS
project, we have de ned a checklist of properties that must be provided [
For systems such as the Open PHACTS Discovery Platform (OPSDP) [
A VoID linkset includes three key pieces of information: the dataset that is the subject of the link triples, the dataset that is the object of the link triples, and the predicate used in the links. Based on this information it is feasible to compute transitive co-reference links based on the properties of the link predicate. For example, given the linksets A p! B and B p! C which link the datasets A and B, and B and C respectively, with the link predicate p, and let p be the predicate owl:sameAs, then it follows through the properties of the link predicate that we have the linkset A p! C which links datasets A and C.
However, as shown by Halpin et al [
The limitation of the linkset predicate approach stems from the generality of the linking predicate and thus the lack of domain knowledge that it conveys. One
Term Justi cation Chemical entity The concepts linked represent the same chemical entity. sio:SIO 010004 Gene The concepts linked are conceptually the same gene. sio:SIO 010035 InChI Key The concepts linked have the same InChI Key. cheminf:CHEMINF 000059 Protein The concepts linked are conceptually the same protein. sio:SIO 010043 Protein coding gene A gene resource and a protein resource are being treated sio:SIO 000985 as being equivalent Table 1: A subset of the vocabulary terms used to capture the justi cation of a linkset and the operational equivalence that is interpreted. sio represents the Semantic Science Integrated Ontology namespace and cheminf the Chemical Information Ontology namespace. approach to overcome this, whilst still retaining the notion of a VoID linkset, would be to mint a new linking predicate for each notion of equivalence; these could be created as sub-properties of existing mapping predicates. However there is a major social barrier to such an approach; gaining consensus on the required linking predicates and updating the existing links in the datasets to use these new link predicates. As such, it is unlikely to gain traction.
Another alternative is to annotate the linkset descriptions with additional contextual data; this enables the use of the existing links unchanged. We term this the justi cation for the linkset; the notion captured is the scienti c interpretation of the operational equivalence applied by the linkset. For example, two chemical datasets, A and B, that are linked because they have the same InChI string would express this relationship in the linkset VoID header with the triples :A-B_Linkset void:linkPredicate skos:exactMatch ;
bdb:linksetJustification cheminf:CHEMINF_000059 .
where :A-B_Linkset is the resource that describes the linkset, the link predicate
is declared to be skos:exactMatch, and the justi cation is speci ed using the
BridgeDb vocabulary namespace (bdb6) with the value taken from the Chemical
Information Ontology namespace (cheminf7). The linkset can be expressed as
A pj! B where the justi cation j is cheminf:CHEMINF 000059 and the link
predicate p is skos:exactMatch. The set of supported justi cations within the
Open PHACTS IMS can be found in [
Based on the justi cation of linksets, we can compute transitive linksets. For example, we can generate a linkset between datasets A and C through some
intermediary dataset B if there is a linkset between A and B and one between B and C such that both linksets have the same justi cation. De nition 1 formally gives the rule for computing transitive linksets based on their linkset justi cation. Note that we do not require that the linksets have the same link predicate. The resulting transitive linkset is given the weaker of the two link predicates with a hierarchy of owl:sameAs skos:exactMatch skos:closeMatch rdfs:seeAlso: Thus, if p was the link predicate owl:sameAs and r the link predicate rdfs:seeAlso, the computed linkset A rj ! C would have the link predicate rdfs:seeAlso.
Given datasets A, B, and C, linksets A pj! B and B rj ! C both with the justi cation j and link predicates p and r respectively then we can generate the linkset { A { A j r ! C if p j p! C if r r;
By iteratively applying the rule given in De nition 1 it is possible to compute chains of linksets that use the same justi cation. However it is possible to enter an in nite cycle; thus the IMS implementation prevents the same linkset being used more than once in a chain. As part of the provenance of the computed linkset, the linksets that are used to compute it are tracked. 4.3
Within the life sciences it is common to use gene names as proxies for protein names since gene names are shorter and more standardised. It is easy for a human to distinguish when this is being done but impossible for a computer to distinguish. A key requirement for the OPSDP is to permit a user to enter with a gene name that is then resolved to a URI for that gene. However, it should be possible to retrieve information about the target { a protein, or group of proteins { for which the entered gene name is a proxy. This means that it must be possible to state that a gene and a protein are operationally equivalent, i.e. to have equivalence across semantic types.
This is straightforward using the linkset justi cation approach. We introduce a new justi cation for a protein coding gene (sio:SIO 000985), see Table 1. The complication comes when computing the transitive linksets to enable a user entered gene to relate to the protein information in each of the datasets. The transitive computation now needs to support equivalence across semantic types with di erent justi cations. In particular we want to support chains of one or more protein linksets, a protein-gene linkset, and one or more gene linksets. It is important to prevent the use of protein-gene linksets to go from a gene to a protein and back again; this is to prevent a chain of links whereby we end up with protein X being claimed to be the same as protein Y due to crossing the semantic type boundary multiple times.
Additional contextual information is required from the linkset description.
Speci cally, the semantic type of the data being linked. This needs to be
captured at the linkset level since datasets can contain multiple semantic types,
e.g. ChEMBL, DrugBank and Ensembl [
De nition 2 extends the co-reference transitive computation rule given in De nition 1 to support cross-type mappings. Note that for simplicity we have omitted the link predicate from the rules in De nition 2. These are derived in the same way as for the rule given in De nition 1. The rst clause of De nition 2 is a combination of the clauses in De nition 1, with the additional constraint that all of the datasets involved are of the same semantic type. The second clause allows for a linkset of the same semantic type and a linkset with a cross semantic type justi cation to be combined, with the resulting linkset being given the cross-type justi cation.
datasets, and 0 be semantic types with 6= 0, and j and j0 be linkset justi cations such that j links datasets with the same type and j0 links datasets across types. Then the following two rules hold for transitive linkset computation: { Same semantic type and justi cation
A j ! B
A ^ B j ! C j ! C { Cross semantic type and justi cation
A j ! B ^ B j0 ! C 0 j0 ! B 0 ^ B 0 j ! C 0 A _ j0
A ! C 0
By iteratively applying De nition 2 it is possible to compute the co-reference of URIs across proteins and genes. Note that the de nition only permits a single cross-type link justi cation, e.g. \protein coding gene", to be used in any chain, although arbitrary numbers of same type justi cations, e.g. \same protein" or \same gene", can be applied. This is due to the consequence of the second rule being given the cross-type link justi cation. Thus, a chain of links resulting in inaccurate mappings is prevented. 5
The Open PHACTS IMS implementation is an extended version of BridgeDb to support cross-references over linked data sources, i.e. supporting the use of URIs to represent records in datasets. The source code is available from https://github.com/openphacts/IdentityMappingService and the service is accessible through the Open PHACTS API, https://dev.openphacts.org/.
The IMS has implemented the transitive co-reference computation rule given in De nition 2. Computed linksets are created with dataset descriptions giving full details of the datasets linked and the justi cation for the link. A screenshot of the web interface to the IMS is shown in Figure 2 with the results for a look-up for the UniProt insulin receptor URI. 5.1
For the 1.3 release of the OPSDP, the IMS was supplied with 104 linksets from
9 providers linking 37 datasets and containing 7,096,712 links. These are shown
in the visualisation in Figure 3a; nodes represent datasets and edges the linksets
between them, with the colour signifying the linkset justi cation. Note that
the large number of linksets is a consequence of splitting links based on their
semantic type { resulting in multiple linksets between some datasets { and in
the case of Ensembl the linksets were further split by species with 12 species
covered. The visualisation highlights the Open PHACTS design decision to use
a small number of datasets as mapping centres; chemical alignment is performed
through the Open PHACTS Chemical Registration Service [
Following the transitive co-reference computation, there are 883 linksets containing a total of 17,383,846 links. These are shown in the visualisation in Figure 3b. The left side of the visualisation is dominated by the linksets that match InChIs via the Open PHACTS Chemical Registration Service while the right side (a) Visualisation showing connectivity of the linksets provided as input to the IMS. (b) Visualisation showing connectivity of the linksets after co-reference computation. Legends
Colour Justi cation URI Red InChI cheminf:CHEMINF 000059 Purple Chemical entity sio:SIO 010004 Green Protein sio:SIO 010043 Orange Gene sio:SIO 010035 Blue Protein coding gene sio:SIO 000985 Light blue Pathway sio:SIO 001107
Yellow Pathway name edam:data 2342 Code Dataset Code Dataset AERS Adverse events reporting system Hm Human metabolite database Bg BioGrid I InterPro Ca Chemical abstracts service Ip International protein index Chebi Chemical entities of biological interest L NCBI Gene Ch.M ChEMBL molecule MGI Mouse genome informatics Ch.T ChEMBL target MSH Medical subject headings Ch.TC ChEMBL target components Pw Pathway ontology Ck KEGG Compound Om Online mendelian inheritance in man Cw ConceptWiki OCS OPS Chemical registration service Cpc PubChem Compound Pd Protein databank Cs ChemSpider Q NCBI Reference Sequence Database D Saccharomyces genome database R Rat genome database Db.D Drugbank drugs S UniProt Db.T Drugbank targets Ug UniGene Em European nucleotide archive Up UniParc En Ensembl Wi Wikipedia F FlyBase Wp Wikipathways Hac HGNC accession number Z Zebra sh information network Fig. 3: Visualisations showing connectivity of the linksets. is dominated by protein coding genes due to the application of the cross-type co-reference rule. The visualisation highlights a second Open PHACTS design decision; to use ConceptWiki8 { Cw in the gure { as the source for text-to-URI translation. This is shown by its connectivity to every other dataset; one of the reasons for computing the co-references. As a consequence of these co-reference computations, users of the OPSDP are able to enter with any chemical, protein, or gene URI known to the IMS and retrieve information about chemicals or targets. 5.2
There are several bene ts to the co-reference computation approach that has been implemented as part of the OPSDP IMS; not least of which is the increased inter-connectivity across the datasets.
First it eases the burden on the dataset providers; with only a small number of additional metadata triples being required to provide justi cations and semantic types. This allows systems such as the OPSDP to exploit the fact that many datasets already contain links within their data to instances in other datasets; particularly those who publish their data as RDF. However it cannot be expected that they can link to every dataset that is required for every possible use case. Thus, by applying the co-reference computation we are able to infer additional connectivity across the data.
Second the co-reference computation tightly controls what can be equated, e.g. only chemical entities can be related through an InChI, and which justications are allowed to cross semantic types, e.g. protein coding gene relating genes and proteins. These safeguards ensure that the result of the co-reference computation matches the expectations of the domain scientists.
User evaluation of the results of the co-reference computation is on-going with domain scientists. The IMS has been successfully deployed by the OPSDP enabling the integration of data across the data sources shown in Figure 3. 6
Identi ers.org [
Bio2RDF [
as a source of mapping information for the IMS and the IMS already returns the Bio2RDF identi er as an alternative URI.
BridgeDb [
Another closely related area of research is focused on generating tools for identifying links between pairs of datasets; either at the schema level Ontology Matching or at the instance level. The latter of these is most relevant to the work in this paper. Since 2009 there has been an instance matching track9 in the annual ontology matching competition10 to compare such tools. The most recent results are available from http://www.instancematching.org/oaei/imei2013/ results.html. The links generated by these instance matching tools could be used as input to the IMS. 7
In this paper we presented the challenges for co-reference across life sciences datasets that stem from each dataset using their own identi er scheme. We have argued that there is not a one size ts all notion of equivalence across concepts in these di erent datasets since they model the data at di erent levels of granularity, e.g. should a drug entry be equated to an entry about the chemical compound. Additionally, users of the data want to apply varying notions of equivalence based on the task they are performing, e.g. should genes and proteins always be equivalent. As such we have proposed that the notion of operational equivalence should be captured in the linksets that relate a pair of datasets as the justi cation for the linkset. The advantage of stating it as a justi cation rather than a mapping predicate is that existing linksets can be easily extended.
Due to the fact that each pair of datasets is not related in the web of data, we have developed rules for transitively computing co-reference across datasets. To support scenarios where genes and proteins should be equated, the co-reference computation allows the crossing of semantic types. We presented a rule for preventing undesired co-references being computed whilst ensuring that concepts within a given type are completely covered. This has been implemented in the
November 2013
10 http://oaei.ontologymatching.org/ accessed November 2013
accessed
Open PHACTS Identity Mapping Service. Details of how the IMS is used within
the OPSDP can be found in [
As future work we will allow applications to apply di erent scienti c lenses
[
The research leading to these results has received support from the Innovative Medicines Initiative Joint Undertaking under grant agreement number 115191, resources of which are composed of nancial contribution from the European Union's Seventh Framework Programme (FP7/2007- 2013) and EFPIA companies' in kind contribution.