Main

The chicken and the egg of scholarly communication

In data-intensive sciences, text is neither the only nor the most effective way to share scientific information. Aware of the paradox, we reintroduce the metaphor of the chicken and the egg to underscore our thesis that there is no meaningful information without data and conversely, data cannot be generated nor valued without prior knowledge. If we assume data to be the eggs, which need brooding (curation) to become chickens (articles), and we require the mating of complementary units of information to generate yet more fertile eggs, we have a reasonable frame of reference.

When datasets were sparse and only connected to the lab that produced them, we would brood every one of them, protect (patent) them and work on them in isolation in order to 'sell' them as chickens, usually in the form of a largely narrative article. Other scientists need to combine a minimum of two existing publications to generate new eggs and breed more chickens. However, chickens have become overabundant: more than 20 million articles exist in biomedicine alone. More recently, valuable aggregations of data were brought online (for example, data sets in GEO, curated databases such as SwissProt and locus-specific human gene variation databases (locus-specific databases such as the Leiden Open Variation Database LOVD). Now, data (eggs) have become a direct source of new in silico discoveries and a unit of scientific trade.

But the scientific market has no way to value eggs because the entire system is built upon judging and exchanging chickens for acknowledgement and credit (through citations and other measures of impact). On the other hand, for effective and evidence-based breeding, we need the eggs as well as information from the parent chickens to assess the value of the eggs. This is where a major challenge lies: in the long overdue adaptation in scholarly communication. The data-intensive science wave that has come over us calls for innovative ways of data sharing, stewardship and valuation. We must respect the connection between the articles and the data and value both appropriately.

A new market for data

We have all heard lamentations about datasets being difficult to find and the painfully slow increase in annotation and curation by the scientific community1. One bottleneck is the lack of a scientific reward system for depositing and curating data outside the mainstream of publishing conventional articles. But how do we implement such a system? In this issue, Giardine et al.2 argue for “microattribution”3, providing credit for contributors and curators of entries in databases of human gene variants. Here we comment on the technical as well as social challenges associated with broadening and sustaining that valuable approach.

Although we argue that the narrative form of scholarly communication will continue to be needed in the future, we also recognize that data-intensive sciences need computer-readable information4. It follows that de novo claims and the supporting data should be exchanged in machine-readable, unambiguous format.' Ideally, this should be created at the same time the descriptive text is composed. Articles should tell us why we should believe the underlying data and the conclusions drawn, and they are perfectly suited for that task as they are.

A graphical analysis of the problem

We propose a new way to represent data, information and, in particular, assertions in the form of nanopublications5. A nanopublication is essentially the smallest unit of publication: a single assertion, associating two concepts by means of a predicate in machine-readable format with proper metadata on provenance and context6. Each concept in a nanopublication has an unambiguous, non-semantic, stable unique and universal identification (UUID), to which different Uniform Resource Indentifiers (URIs) can be resolved5. Nanopublications support in silico knowledge discovery, tapping massive treasures of implicit information7,8.

As an illustration of our vision, we represent the current state of scholarly communication in a graph of visualized nanopublications (Fig. 1). In the caption, we show that this picture can also be represented in narrative text, which is much more easily understood by people than the picture. However, making the picture did help structure the argument, making the text easier to write, and the picture is 'reasonable' for computers. However, it is hard to go the other way around, that is, to reconstruct the picture from the text.

Figure 1: The current state of scholarly communication.
figure 1

Red ovals represent points where the current system is 'broken'. All predicates are represented as small ovals. An asterisk (yellow star in the figure) represents 'conditionality'. Illustrated concepts (large ovals) are discussed below in bold. Data-intensive science produces large datasets. It also produces articles in which the narrative of the experiments, methods and conclusions and some figures and tables are presented. The datasets that are too large to be part of the article itself are currently 'linked' as 'supplementary data'. Unless (as indicated by stars) there is good data stewardship by the journal (and journal-governed storage), the link to these data is a weak point in the system. Articles have a proper ID and can therefore be cited, and citation metrics are possible. However, datasets and databases as well as figures and tables usually (as indicated by stars) do not have such a stable, unambiguous identifier. Neither do terms, thus, retrospectively mapping terms in the narrative, in tables and in figures to unambiguous concept identifiers is extremely hard, leaving these research objects suboptimal for text mining and data mining. Therefore, it becomes difficult to 'recover' original claims from these research objects, and they cannot be properly cited. With text-based articles as the primary focus of scholarly communication, data sharing (that is, deposition and sharing) and community annotation will remain invisible for mainstream citation metrics and will not accrue the proper credit or reward.

We needed 217 words to describe the 21 assertions in Figure 1. Please note how we introduced near-synonyms, like 'scientific awards' for 'professional awards'. By using ambiguous terms and complicated sentence structure, we all contribute to 'knowledge burying'4. Above all these difficulties hovers another problem: much of what is worth mining is simply not findable or accessible with the current query methods and firewalls.

Imagine that we published Figure 1 as a set of properly interconnected machine-readable nanopublications. This picture would then indeed be perfectly 'reasonable' in semantic computation engines such as the LarKC system7. Computer reasoning would most likely shift the narrative articles oval in Figure 1 from its central position slightly off to the side and insert nanopublications in the central position (Fig. 2). This move would make all of the problematic 'red predicates' in Figure 1 disappear by the virtues of the machine-readability of nanopublications. We think that it is worth looking at that suggestion.

Figure 2: A proposal for the future of scholarly communication.
figure 2

The concepts and predicates are represented as in Figure 1. By placing machine-readable nanopublications at the core of communication and moving the narrative slightly in the graph, many problems may be solved. First, datasets, tables and the underlying data of figures (graphs), as well as their captions, can be represented as nanopublications. Because all concepts in a nanopublication graph, including those in the provenance and context parts, have a proper UUID, they automatically link through the provenance to their source and their position in that source. The narrative article now becomes supplementary to the data: it provides the detailed description of the context as well as the rhetoric building the argument as to why the data are valid and the claims correct. Good news for article publishers is that each nanopublication in a conventional article, a supplementary dataset (for instance, for two co-expressing genes) or a related database (for instance, for a variant-phenotype association) is now intrinsically findable, citable and hard-linked (through the article UUID in the provenance) to the underpinning article. Therefore, nanopublications will have the potential to increase the hit rate of an article and will promote proper citation of the article in which the 'first' claim was made. As claims are now individually citable, citation metrics are possible, and community curation and annotation can be traced back to the contributing scientists. Now, proper scientific reward is possible by precise attribution, and in the case of the simplest multiples, nanoattribution.

Some argue that the rhetoric in articles is difficult to mine and to represent in the machine-readable format. Agreed, but frankly, why should we try? All nanopublications will be linked to their supporting article by its DOI. Many conceptually unique biological assertions are repeated time and again in texts and databases. Capturing the majority of them using a variety of trusted sources is one way to collect almost all relevant biomedical assertions ever made and to enrich them with a dynamic evidence factor based on frequency and conditionality6.

When reasoning over the associations represented in such a computer readable set of assertions a scientist may have reason to check—even to doubt—any particular assertion, in the graph. Ideally, the list of underpinning articles and other data sources supporting that claim should be just a click away, enabled by the nanopublication provenance. The researcher can now judge the validity of the claim in question much better by reading the articles than by trying to judge a rhetoric argument that is painfully distorted into machine-readable format.

Practical implementation for knowledge discovery

A system very close to the one described in Figure 2 will soon be put to the test in the recently launched Open PHACTS project of the Innovative Medicines Initiative to create an Open Pharmacological Space. This project will be based on representation technologies and tools currently used to expose a wide range of data in machine processable forms9.

Giardine et al.2 review a number of database entries relating human gene variants to hemoglobin phenotypes. The technical feasibility of microattribution for the submitted variant-phenotype associations has been the subject of a pilot study. We have collaborated to mine nanopublications from the article text as well as from the underlying databases and supplementary data. The pilot results illustrate the points made above. Using text mining and manual inspection, we recovered 698 nanopublications from the narrative covering all biomedical concepts in the paper. Only 13 of these directly assert genomic variation, for example, those of the composition [HGVS gene variant name][has][variant frequency].

Using simple unambiguous parsing routines, we have represented (in Supplementary Tables 1 and 2, respectively) two classes of nanopublications in the supplementary tables of Giardine et al.2 of the form [HGVS gene variant name][has][variant frequency] and [HGVS gene variant name][has][OMIM allelic variant ID]. For those two classes, we found 1,855 instances of nanopublications now deposited at http://www.nanopub.org/ in XML format. These open access nanopublications are now provenance-linked to the article and the databases (and thus to the submitters, curators and authors). The provenance-linked nanopublications have the potential to increase the conventional citation rate of the article and database as well as provide the potential for microattribution. We estimated that the databases described in the article2 currently contain around 40,000 nanopublications with meaning to geneticists and with citation potential.

Importantly, in this article, we deal with a particularly vulnerable subset of 'concepts', namely the so-called 'variants' in gene sequence. The HGVS nomenclature of such variants may increasingly be enforced by some prescient journals10, but trying to find these variants and their synonyms in the broader literature is a notoriously difficult task that can only be done with some degree of success if one has access to the full text and, more importantly, all the supplementary data. In a related text-mining analysis, out of 4,940 different variants of 11 genes from the LOVD, only 16 variants could be identified in 10 million PubMed abstracts. Again, we see the tremendous advantage of data publication over text mining in exposing potential nanopublications. These results indicate that authors should construct and publish their data as nanopublications in tabular, ID-based databases as part of their submission and support these tables with narrative text.

The brooding question: can data curation duties be traded?

Only a minority of nanopublications in databases and datasets will ever make it into a narrative as an explicit textual assertion. Even if they do, they will be very difficult to recover retrospectively, for reasons related to access and the failings of mining technology, in confronting ambiguity and sentence construction. We estimated that describing the supplementary data of Giardine et al.2 would require roughly 4 million words, with the result being a corpus hardly readable by machines.

On the other hand, a single LOVD website (http://www.dmd.nl/) consistently enjoyed more than 50 citations annually over the past three years. It is therefore reasonable to assume that proper formatting and exposure of the nanopublications contained in diverse sources such as locus-specific databases could allow these resources to be recognized for the important scientific contributions they actually are. Appropriate standards for proper measurement of these citable items seem to be the only remaining obstacle. So, let us agree to evolve these and to communicate more effectively.

URLs. Gene Expression Omnibus (GEO), http://www.ncbi.nlm.nih.gov/geo/; SwissProt at ExPASy Proteomics Server, http://expasy.org/sprot/; Leiden Open Variation Database (LOVD), http://www.lovd.nl/2.0/; The IMI Open PHACTS project, http://www.openphacts.org; PubMed, http://www.ncbi.nlm.nih.gov/pubmed/; HbVar, http://globin.cse.psu.edu/hbvar/; OMIM, http://www.ncbi.nlm.nih.gov/omim.