Biomedical researchers consume and analyze PubMed abstracts on a daily basis seeking to update their existing knowledge with insights from newly published literature. Plain text descriptions fail to deliver contextual knowledge to users who require a comprehensive understanding of the content of a paper before deciding to access it. To achieve this biological named entities described in the abstracts must be linked to their related entries in biological databases and established controlled vocabularies such as SwissProt and Gene Ontology. Semantic Assistants support users in content retrieval, analysis, and development, by o ering context-sensitive NLP services directly integrated in standard desktop clients, like a word processor. They are implemented through an open service-oriented architecture, using Semantic Web ontologies and W3C Web Services. Here we present a deployment of the Semantic Assistants framework to provide links from mutation, protein, protein property, gene and organism mentions in abstracts to their related entry in standardized biological databases and controlled vocabularies. The underlying text mining pipeline used to identify named entities has previously shown high levels of precision and we make this functionality easily accessible through a Semantic Assistant, to end users when reviewing PubMed abstracts in through a Firefox client.
The proliferation of annotation services for biologists reviewing scienti c literature is based on the
adoption of web services designed to provide links to contextual information provided in online databases.
These annotation services, such as Re ect [1], tag gene, protein and small-molecule names and link them to
external resources with sequence and structure information for given default organisms. Further
deployments of annotation services, such as BioDEAL [
Recent work on Semantic Assistants [
In previous work [
Reported semantic assistant deployments rely on clients that are plug-ins for text editors as e.g. OpenO ce. In current work we have implemented a Semantic assistant client with the use of Greasemonkey [5], an extension to Mozilla Firefox allowing users to install scripts supporting augmented browsing, i.e. making changes to webpage content. The script we have built pre-processes a PubMed entry page and sends the abstract, via a Java Servlet wrapping the Semantic Assistant Java API, to our mutation impact extraction service and annotates the abstract with the results as shown in Figure 1. An important di erence between text in word processors and web page content is that the latter is not editable and contains both texts of interest and, depending on the web page selected, content that can be considered nonsense. For a given page the list of web services have to be extended according to the content they will process (which webpage and which part of that webpage). As an example, the NLP service "mutation impact extractor" can be customized to a "mutation impact extractor for pubmed abstracts". Moreover text processing algorithms for some NLP services may be context dependent in that they may rely on co-occurrence of terms within a certain distance from named entities. Consequently multiple texts or other surrounding content might disrupt the underlying algorithms for entity recognition and disambiguation. This makes the construction of semantic assistant clients for web browsers a bit more tedious. We propose a client that will support browsing, augmented with semantic tags, by using Ajax technologies in combination with the existing semantic assistant architecture. The core of the client is a Java servlet making use of a Java API containing a precompiled client-side abstraction layer which performs the actual communication with the server. The exterior of the clients consist of a set of scripts on which they can make changes to HTML pages with annotations retrieved via asynchronous calls to the Java servlet.
The backbone of a Semantic Assistant is the NLP service that process text and outputs annotations or
deduced facts. We have previously developed such an NLP algorithms and semantic infrastructure which
nds, annotates and disambiguates biological entities by using a combination of gazetteer-based approaches
and methods for relation detection. The following entities are currently extracted: proteins, genes,
organisms, point mutations, protein properties and mutational impacts on protein properties. In particular
these algorithms facilitate the grounding (linking) of proteins and mutations to SwissProt entries and
correct position on amino acid sequences [
Firstly the cross-linking entities found in text with their real-world counterparts, called grounding, requires
the extraction and normalization of mutation mentions, for which we used the MutationFinder system [
Secondly the extraction of mutation impacts from documents relies on both the identi cation of protein
functions, found in noun phrases and the extraction of directionality terms also found in sentences adjacent
to mutation mentions. This is achieved as the result of identifying Gene Ontology Molecular Function
terms involving activity, binding, a nity or speci city as the head noun in noun phrases identi ed using
the multi-lingual noun phrase extractor [
Although we can show that our NLP services work for PubMed abstract, we know that the performance of
the underlying grounding algorithms is decreased when switching from full-text to abstracts only. To
enhance this performance and to retrieve more information, the semantic assistant client should send the
full-text content to the service, this can be made via existing web services like EFetch [
For future work, we will host a web site with listings of available services. The users can, when choosing a service, also restrict it to only be run on certain web sites as e.g.: http://www.ncbi.nlm.nih.gov/pubmed/* To ensure that new customized services can be constructed and consumed by both service providers and end-users, template client scripts should be made available where parameters can be set for content restriction, i.e. only send content in speci c document elements as e.g.: <p id="abstract">content</p> And also to set custom styles for output annotations. These customized client scripts can then be stored and listed as services, available to other users as well.
The adoption of new frameworks providing online annotations to already published content is an emerging theme trend in life science knowledge discovery. In this brief work we have shown that existing algorithms for grounding of mutation mentions and related content can be deployed to great e ect albeit in prototype scale application. This serves both as a strong motivation for deploying further semantic assistants for other named entities and as a test bed to solicit further requirements from end users. In the process of deploying this prototype in an open source web browser we have identi ed constraints that will impact the design of next generation Semantic Assistants.
NLP: Natural Language Processing; OWL-DL: A Web Ontology sub-Language; API: Application Programming Interface; HTML: HyperText Markup Language;
The authors declare that they have no competing interests. References 1. Pa lis E, O'Donoghue SI, Jensen LJ, Horn H, Kuhn M, Brown NP, Schneider R: Re ect: augmented browsing for the life scientist. Nature Biotechnology 2009, 27:508{510. 5. Greasemonkey. [http://www.greasespot.net].
The abstract is tagged according to the output of our mutation impact NLP service. Here we can see that the protein Haloalkane Dehalogenase is tagged together with two mutations on the same protein. Algorithms for protein and mutation grounding used by the back-end NLP system make sure the mutations refer to the correct position on the correct amino acid sequence as seen in the popup box, produced when the mouse pointer hover over a tag surrounding the F294A point mutation mention.