On the Web, it is possible for anyone to publish linked open data as RDF. Whilst this has huge potential to bene t data integration e orts, it highlights challenges of assessing data quality and trust. Nanopublication is an approach to data and knowledge publication in which assertions are individually encoded in RDF along with details about provenance, context and attribution. Collectively these details form a body of evidence for (or against) an assertion, which can be used as quality and trust criteria during data integration. In this position paper, we highlight the features of the Nanopublication speci cation that can be used as quality and trust criteria for life science data. We introduce the concept of cardinal assertions; assertions that are derived from the aggregation of multiple nanopublications to give an evidence value. We also identify a role for cardinal assertions in the evolution of evidence over time, supporting the re-evaluation of data and hypotheses.
As the corpus of life science knowledge grows, along with the increasing amount
of structured and unstructured life science data available on the Web, one of
the challenges faced by life science researchers is the evaluation of evidence for
biological assertions [
Data integration is not only technically challenging, but the result is also subjective, as data sources are curated to di erent standards and integrators have di erent motivations. In a scienti c context, where quality and trust are paramount [9], it is important to be able to discover what has been integrated, and why. Annotations of assertions, and the methods used to derive them, are important factors in deciding their scienti c quality. In many databases, bibliographic references are associated with assertions, which gives a researcher a broad indication of where they can nd out more about their provenance. The importance of increasing the resolution of evidence for an assertion is apparent in the use of evidence codes by curated databases, such as those used for Gene Ontology annotations [10] or by the BioGRID Interaction Database [11].
In this position paper we consider how evidence is currently used to support scienti c assertions in the life sciences, and the potential impact that the structured provenance for assertions provided by the Nanopublication speci cation might have on the way research data is published, used and integrated. 2
A lack of fundamental information about biological assertions, such as their source and the date that they were last updated, can make it di cult to assess their quality [12]. Nanopublications 4 have been proposed as a way to encode and publish individual assertions using Semantic Web and Linked Data principles [13, 14, 15]. Authorship and timestamp metadata are mandatory components of Nanopublications, ensuring a basic level of trust for any Nanopublication. Each Nanopublication contains exactly one assertion, which may be encoded as one or more RDF triples in a named graph; the assertion graph. Further provenance of the assertion is speci ed as annotations of the assertion graph in a second named graph; the provenance graph. Figure 1 shows components of a Nanopublication that might be published by a database containing protein-protein interactions 5. In this case a reference to a publication and a laboratory technique, elements that are commonly found as supporting evidence, are encoded as provenance. Other types of assertion such as functional annotations of genes or drug-target associations can be represented just as easily in the Nanopublication framework. With the structure of Nanopublications, it is possible to list the evidence for an assertion by querying for the provenance of the assertion across known Nanopublications. Simply listing all of the associated evidence is a lightweight approach, but it can be used to make an assessment of the level of evidence available. Like with evidence codes, researchers and application developers will be able to make broad but valuable distinctions between assertions that have no evidence, are 4 Fine details of Nanopublication structure and content are yet to go through a full standardization process, so we only highlight broadly agreed principles here. 5 The example is adapted from an entry in the BioGRID database predictions, author statements, and assertions that have been derived experimentally. In the context of RDF warehouses, the consistent representation of evidence in provenance graphs enables researchers to include queries that reveal the level of support for interesting connections. 3
So far we have considered how Nanopublications can provide a means to collect and integrate assertions published on the Web. Beyond integration, evidence is also often aggregated computationally such that one value of overall con dence is produced for an assertion. An example of evidence aggregation is provided by StringDB [16], which contains data about associations between proteins. For each type of evidence, a probabilistic score indicates how likely a functional association is considered to be. In terms of Nanopublications, these associations could be published as assertions that are supported by the type of evidence and con dence score as part of the provenance graph. In addition, a combined score is calculated that takes all of the evidence into account. This kind of combination and evaluation of evidence from di erent sources is not unique to StringDB. Another example of aggregation comes from neXtProt [17], where evidence is assigned manually by a Gold, Silver or Bronze rating. These practices have prompted us to consider how such aggregated evidence ts into the Nanopublication framework.
Here we de ne a cardinal assertion. Cardinal assertions are the result of aggregating the evidence associated with identical assertions to produce a new measure that represents a consensus of that evidence (Figure 2). A cardinal assertion comes with two speci c criteria: 1) it links to the Nanopublications containing the source assertions 2) the method of generating the overall con dence score should be clearly de ned and linked to. The links should be encoded in a provenance graph and published in the Nanopublication format. These criteria ensure that the component assertions and their provenance can be identi ed at a later date, and that the aggregation method can be evaluated and repeated. As cardinal assertions use the Nanopublication approach, they also inherit the ability to be published as citable entities for which the aggregator can get direct attribution when they are used or referenced. A set of cardinal assertions represents the judgement of a data aggregator as to what available evidence should contribute to the state of the art of knowledge and by how much. For instance, StringDB could aggregate evidence from Nanopublications about interactions and expose its combined evidence scores as Cardinal assertions in RDF. Through the creation of a set of cardinal assertions, a data aggregator removes the need to repeatedly evaluate all of the Nanopublications in a set. 4
Future Perspective I: Curation of Cardinal Assertions With open access to Nanopublications, researchers have the opportunity to collect and aggregate them into Cardinal assertions to summarize evidence. An authoritative collection of cardinal assertions could represent a valuable, structured interpretation of the state-of-the-art of knowledge in a particular domain of the life sciences. Such data sources could form the foundation of a new level of trust and quality of data used to form new hypotheses and analyze data that is more in line with that expected from the scienti c method. Here we consider some of the aspects of managing Nanopublications and cardinal assertions that an authoritative data provider might need to consider.
A data aggregator will be able to curate nanopublications as trustworthy and reject others based on their provenance, for example where provenance is insufciently described, or indicates that an assertion is derived from a methodology that has later been identi ed as unreliable. Di erent aggregators may have di erent parameters for what constitutes trustworthy evidence, for example, whether text-mining predictions are included. The ability to make these decisions is provided by the content of the Nanopublication provenance graph.
If a data aggregator chooses not to include a particular (type of) Nanopublication in their dataset, they also have the opportunity to publish this fact as further data so that 1) others can see that a particular Nanopublication has been rejected or not considered by them and 2) the reason for its absence is clear. Similarly, accepting a particular Nanopublication during the curation process represents an endorsement of its quality. For consumers of the data it is important to be able to establish which nanopublications were or were not accepted by the data provider, and why, so this should be a queryable part of the structured data. From this management perspective, we are working on a system that can keep track of which nanopublications have been accepted and which have not. We intend to integrate this functionality with an RDF triplestore to transparently add metadata about this process. 5
Life science knowledge is constantly evolving as new experiments are performed,
new data are produced and our ability to more accurately describe existing
knowledge improves. This evolution is increasingly characterized by evidence
derived from high-throughput experimentation and bioinformatics methods [18].
For example, it is reported that some ninety-eight percent of all Gene Ontology
annotations are uncurated and inferred through in silico processes [
A second example of knowledge evolution comes from the interpretation of high-dimensional datasets. For example, in 1998 Spellman et al. claimed to have elucidated the set of cell-cycle regulated genes of the yeast Saccharomyces cerevisiae in a landmark gene expression paper [20]. Alternative transcriptomic experiments have subsequently implicated di erent, but overlapping sets of genes. In addition, di erent data analysis methods produce di erent sets of genes from the same experimental data [21] leading to uncertainty about the meaning of the data [22].
If assertions like those described above were published as Nanopublications over time, a data aggregator can use the provenance information to recalculate the evidence values for their assertions and then publish them as Cardinal assertions. As evidence values change over time, it will in uence the interpretation and use of those of cardinal assertions, which can be very signi cant if they were used in a computational analysis pipeline e.g. [23].
Nanopublications themselves have been proposed as immutable things; once they are published their content should not change. This is also true for cardinal assertions, in that as new evidence is taken into consideration, new editions of cardinal assertions should be generated. This will create a chain of assertions that represent the state of the overall evidence for a cardinal assertion over time. To be able to compare between editions through querying it is important that each can be accessed or reconstructed, and that the nanopublications that a ected the evidence value are clear.
As life science datasets are often curated to di erent standards, and descriptions of quality are sparse, integrators and users of integrated RDF datasets would have trouble measuring the quality of assertions in an aggregated dataset. Subsequently it would be di cult to query or inference over them with any degree of trust for hypothesis generation or data analysis. Poorly supported, erroneous or obsolete assertions would be di cult to identify. The Nanopublication speci cation was in part designed with this issue in mind, enabling trust through clear provenance statements for assertions. By introducing Cardinal assertions we hope to extend the functionality of Nanopublications for data providers and consumers who want to generate and use combined evidence. We acknowledge that Nanopublications are only a piece of a solution as the evaluation of scienti c assertions can require deep insights into data provenance that may not be encoded with recognized standards by producers and providers of data. However, in this position paper we have shown potential applications of Nanopublications and Cardinal assertions to reinforce quality and trust aspects of shared life science data. [9] Gamble, M., Goble, C.: Quality, trust, and utility of scienti c data on the web: Towards a joint model. In: Proceedings of the ACM WebSci'11. Volume Proceedings of the ACM WebSci'11., Koblenz, Germany (June 2011) 1{8 [10] du Plessis, L., kunca, N., Dessimoz, C.: The what, where, how and why of gene ontologya primer for bioinformaticians. Brie ngs in Bioinformatics 12(6) (November 2011) 723{735 PMID: 21330331 PMCID: PMC3220872. [11] Stark, C., Breitkreutz, B.J., Chatr-aryamontri, A., Boucher, L., Oughtred, R., Livstone, M.S., Nixon, J., Van Auken, K., Wang, X., Shi, X., Reguly, T., Rust, J.M., Winter, A., Dolinski, K., Tyers, M.: The BioGRID interaction database: 2011 update. Nucleic Acids Research 39(Database issue) (January 2011) D698{ D704 PMID: 21071413 PMCID: PMC3013707. [12] Buza, T.J., McCarthy, F.M., Wang, N., Bridges, S.M., Burgess, S.C.: Gene ontology annotation quality analysis in model eukaryotes. Nucleic acids research 36(2) (February 2008) e12 PMID: 18187504. [13] Mons, B., Velterop, J.: Nano-publication in the e-science era. In: Proceedings of the Workshop on Semantic Web Applications in Scienti c Discourse, Washington DC, USA (October 2009) [14] Groth, P., Gibson, A., Velterop, J.: The anatomy of a nanopublication. Information Services and Use 30(1) (January 2010) 51{56 [15] Mons, B., Haagen, H.v., Chichester, C., Hoen, P.B.t., Dunnen, J.T.d., Ommen, G.v., Mulligen, E.v., Singh, B., Hooft, R., Roos, M., Hammond, J., Kiesel, B., Giardine, B., Velterop, J., Groth, P., Schultes, E.: The value of data. Nature Genetics 43(4) (2011) 281{283 [16] Szklarczyk, D., Franceschini, A., Kuhn, M., Simonovic, M., Roth, A., Minguez, P., Doerks, T., Stark, M., Muller, J., Bork, P., Jensen, L.J., von Mering, C.: The STRING database in 2011: functional interaction networks of proteins, globally integrated and scored. Nucleic acids research 39(Database issue) (January 2011) D561{568 PMID: 21045058. [17] Lane, L., Argoud-Puy, G., Britan, A., Cusin, I., Duek, P.D., Evalet, O., Gateau, A., Gaudet, P., Gleizes, A., Masselot, A., Zwahlen, C., Bairoch, A.: neXtProt: a knowledge platform for human proteins. Nucleic Acids Research 40(D1) (December 2011) D76{D83 [18] Kell, D.B., Oliver, S.G.: Here is the evidence, now what is the hypothesis? the complementary roles of inductive and hypothesis-driven science in the post-genomic era. BioEssays: news and reviews in molecular, cellular and developmental biology 26(1) (January 2004) 99{105 PMID: 14696046. [19] Rhee, S.Y., Wood, V., Dolinski, K., Draghici, S.: Use and misuse of the gene ontology annotations. Nature Reviews Genetics 9(7) (July 2008) 509{515 [20] Spellman, P.T., Sherlock, G., Zhang, M.Q., Iyer, V.R., Anders, K., Eisen, M.B., Brown, P.O., Botstein, D., Futcher, B.: Comprehensive identi cation of cell cycleregulated genes of the yeast saccharomyces cerevisiae by microarray hybridization.
Molecular biology of the cell 9(12) (December 1998) 3273{3297 PMID: 9843569. [21] Lichtenberg, U.d., Jensen, L.J., Fausbll, A., Jensen, T.S., Bork, P., Brunak, S.: Comparison of computational methods for the identi cation of cell cycle-regulated genes. Bioinformatics 21(7) (April 2005) 1164{1171 [22] Futschik, M.E., Herzel, H.: Are we overestimating the number of cell-cycling genes? the impact of background models on time-series analysis. Bioinformatics 24(8) (April 2008) 1063{1069 [23] van den Berg, B.H.J., McCarthy, F.M., Lamont, S.J., Burgess, S.C.: Reannotation is an essential step in systems biology modeling of functional genomics data. PloS one 5(5) (2010) e10642 PMID: 20498845.