Applying linked data approaches to pharmacology: Architectural decisions and implementation by Alasdair J.G. Gray, et. al.
The discovery of new medicines requires pharmacologists to interact with a number of information sources ranging from tabular data to scientific papers, and other specialized formats. In this application report, we describe a linked data platform for integrating multiple pharmacology datasets that form the basis for several drug discovery applications. The functionality offered by the platform has been drawn from a collection of prioritised drug discovery business questions created as part of the Open PHACTS project, a collaboration of research institutions and major pharmaceutical companies. We describe the architecture of the platform focusing on seven design decisions that drove its development with the aim of informing others developing similar software in this or other domains. The utility of the platform is demonstrated by the variety of drug discovery applications being built to access the integrated data.
An alpha version of the OPS platform is currently available to the Open PHACTS consortium and a first public release will be made in late 2012, see http://www.openphacts.org/ for details.
The paper acknowledges that present database entries lack semantics.
A further challenge is the lack of semantics associated with links in traditional database entries. For example, the entry in UniProt for the protein “kinase C alpha type homo sapien” 4 contains a link to the Enzyme database record 5, which has complementary data about the same protein and thus the identifiers can be considered as being equivalent. One approach to resolve this, proposed by Identifiers.org, is to provide a URI for the concept which contains links to the database records about the concept . However, the UniProt entry also contains a link to the DrugBank compound “Phosphatidylserine” 6. Clearly, these concepts are not identical as one is a protein and the other a chemical compound. The link in this case is representative of some interaction between the compound and the protein, but this is left to a human to interpret. Thus, for successful data integration one must devise strategies that address such inconsistencies within the existing data.
I would have said databases lack properties to identify the subjects in question but there is little difference in the outcome of our respective positions, i.e., we need more semantics to make robust use of existing data.
Perhaps even more importantly, the paper treats “equality” as context dependent:
Equality is context dependent
Datasets often provide links to equivalent concepts in other datasets. These result in a profusion of “equivalent” identifiers for a concept. Identifiers.org provide a single identifier that links to all the underlying equivalent dataset records for a concept. However, this constrains the system to a single view of the data, albeit an important one.
A novel approach to instance level links between the datasets is used in the OPS platform. Scientists care about the types of links between entities: different scientists will accept concepts being linked in different ways and for different tasks they are willing to accept different forms of relationships. For example, when trying to find the targets that a particular compound interacts with, some data sources may have created mappings to gene rather than protein identifiers: in such instances it may be acceptable to users to treat gene and protein IDs as being in some sense equivalent. However, in other situations this may not be acceptable and the OPS platform needs to allow for this dynamic equivalence within a scientific context. As a consequence, rather than hard coding the links into the datasets, the OPS platform defers the instance level links to be resolved during query execution by the Identity Mapping Service (IMS). Thus, by changing the set of dataset links used to execute the query, different interpretations over the data can be provided.
Opaque mappings between datasets, i.e., mappings that don’t assign properties to source, target and then say what properties or conditions must be met for the mapping to be vaild, are of little use. Rely on opaque mappings at your own risk.
On the other hand, I fully agree that equality is context dependent and the choice of the criteria for equivalence should be left up to users. I suppose in that sense if users wanted to rely on opaque mappings, that would be their choice.
While an exciting paper, it is discussing architectural decisions and so we are not at the point of debating these issues in detail. It promises to be an exciting discussion!