Biological databases and the biological literature have traditionally been distinct resources. The main difference being that databases are providing structured information, while articles are essentially free text and unstructured. As we have argued in the past 1 2 3 , this is an artificial distinction, in part defined by the way each resource is perceived, and in part because databases are an on-line medium and journals have traditionally been hardcopy. As the scientific literature has moved on-line, the distinction has blurred. Databases have become more like the literature by including an increased amount of annotation, often extracted from the literature 4 . Conversely, the literature has become more like databases by including an increasing amount of supplemental information, including the data derived from the experiments described in the research article.

The blurring of the distinction between the biological literature and biological databases is furthered by open access that (depending on the specific open access license) provides literature in a free and unrestricted XML marked-up form as defined by the National Library of Medicine (NLM) Document Type Definition (DTD). By parsing the literature available in this XML form it is possible to extract semantic meaning. We have taken a relatively simple approach to this by extracting semantics associated with database identifiers and ontology terms 5 . With these terms, which are typically already well defined in a variety of biological databases, it is possible to create interesting associations between database and literature content.

We illustrate this here by integrating the content of PubMed Central (PMC) with that of the RCSB Protein Data Bank (PDB) 6 via the BioLit 5 resource. At this time only about 21% of structures in the PDB (as defined by their PDB identifiers) are referenced in the full text of open access articles contained in PMC, but already some interesting associations can be made. The immediate impact when viewed from the RCSB PDB is that new literature references to the structure are uncovered, even when the primary citation to the structure (if one exists) is not cited in that literature. As described subsequently, appropriate components of that literature can then be integrated with database content and presented as a unified view; a small step towards true literature-data integration.

Many groups have made significant contributions in extracting semantic data from both open- and closed-access literature in the life sciences, although none focus on PDB IDs. In particular, GoPubMed 7 , SEGOPubmed 8 , and Textpresso 9 , all web resources, have improved the classification and searchability of articles by inferring semantic relationships in articles using existing or customized ontologies 9 10 . In structural biology, PDBsum 11 has gained permission from publishers to use selected figures and captions for the PDBsum pages and the figures are extracted from the journal's websites using custom made scripts. The FEBS Letters engaged a collaboration with the MINT database aimed at integrating each manuscript with a structured summary, which precisely reports the protein interactions reported in the manuscript. This is achieved using database identifiers and predefined controlled vocabularies 12 .

We proceed by describing the methods used to establish RCSB PDB-BioLit integration, followed by examples of how that integration is presented and conclude with future directions afforded by free and unrestricted access to database and literature content.

At present research articles for approximately 21% of the structures in the Protein Data Bank (PDB) are found in PubMed Central (PMC). This number is expected to increase rapidly given that many of the world's research funding agencies have specified that the publications associated with the research they fund must be made open access and hence available through PMC. For example the NIH stipulates that research publications they fund be accessible within one year of publication http://publicaccess.nih.gov/.

One of the difficulties when data mining PDB IDs is that at 4-characters in length they may not be unique. For example, the PDB ID 3DNA had several false positive hits that were found to refer to a software package and a wet lab kit, both sharing the name "3DNA." By applying contextual criteria within the text mining and data analysis pipeline such misleading results can be filtered. Still subsequent manual review is the only method for achieving total accuracy. We are maintaining a list of PDB IDs that are known to contain false positive hits and for which stricter criteria are applied. This simple approach has turned out to work well to filter out false positive articles.

The future calls for unique identifiers. Digital Object Identifiers (DOIs) are already provided for each PDB structure, in a manner similar to research articles, so the authoritative reference to the content associated with the DOI can be resolved in an Internet environment. This is a positive step, but at present few research articles provide DOIs to reference structures. Moreover, this does not resolve the finer parts of a macromolecular structure, for example individual polypeptide chains and ligands, each of which are often referenced specifically in research articles.

We describe an initial step in database and literature integration using common and reliable semantics to create the linkage between the two previously disparate resources. The intent is to show the promise that this approach has to providing improved comprehension, not just to users of the RCSB PDB, but if implemented by other databases, to those as well. Semantic association through database identifiers is just a first step in what is possible as PMC increases in content. Ontological terms commonly used by databases can be located in the literature and richer and more contextual interrelationships are possible. In the case of the RCSB PDB a sign of success would be for the user to identify from literature integration a function of the protein previously unknown to them. Tool development to hopefully facilitate this type of discovery is on-going.

AP wrote the BioLit client library, provided the new literature view and drafted the manuscript. MAM & JLF developed BioLit and provided the Web Services. MAM maintains updates of PMC and implemented the false positive filters. BTY and DD provide development and support of the servers. BB contributed to the development of the literature view. PR participated in the design and coordination. PEB conceived of the idea, directed the research and wrote sections of the manuscript. JLF coordinated and managed the BioLit project and helped draft the manuscript. All authors read and approved the final manuscript.