Gene Indexing in the MGI Biocuration Workflow. During primary and secondary literature selection (triage), MGI curators review scientific journals in PDF or HTML format for information relevant to the mouse model organism. We initially screen articles by searching for the terms: Mouse, Mice and Murine. Relevant articles are retrieved in PDF format, cataloged based on content, and assigned an internal reference number, which is entered into a master bibliography. Selected articles are then further categorized and indexed by curators, who identify the type of mouse data contained in the article and tag articles for indexing in the MGI database. Individual curation teams are responsible for managing GO, gene expression, sequence, mapping, phenotype and tumor biology data. Each curation team has its own methodology for indexing, which is our internal process for associating article reference numbers to at least one entity within the MGI database. For the GO curation team, this entity is a gene, typically identified by a gene symbol, name or synonym. During gene indexing, curators identify the primary mouse genes studied in the article. Mouse gene references must first be distinguished from human genes of the same name and from gene mentions associated with transgenic mouse models, which are not incorporated in the GO annotations. If an identified mouse gene is in MGI, the article reference number is associated with the gene in the database and the article enters the curation queue. If the gene is not in MGI, it is identified as a new gene, forwarded to Nomenclature for name assignment, where the new gene is added to the database. The article is then associated with the newly created gene symbol and ready for curation. The gene indexing process identifies papers for further curation of more detailed data. Each paper must be indexed to at least one gene entity before it can be assigned to curators for annotation.

OBA gene mention annotation statistics. In BioPortal Release 2.1, the NCBO Open Biomedical Annotator web service provides visually appealing annotation tag clouds and annotation statistics for plain text submissions processed using the MGI mouse and human gene dictionary (JAXMGD) and a semantic type developed for this dictionary (T998). This screen shot shows the annotation results for the title and abstract of a Journal of Immunology article. The annotation tag cloud correctly identifies Tlr4 and Ly96 as the most prominent genes with the highest relevance scores (scores appear when you mouse over the gene name) discussed in the article. Relevance scores are calculated based on the sum of weights given to each annotation based on the annotation context.

Visualization of OnTheFly and Reflect tagging. OnTheFly enables users to load PDF files for tagging by the EMBL Reflect server. The color coding choices make the marked up document easy to read, however the underlying PDF converter has difficulty translating special characters (such as the Greek letter α in the title) and kerned letters, such as the ‘fi’ in purified and ‘fl’ in inflammation. (A) Tagged entities in this Journal of Immunology article are color coded by type of bioentity and hyperlinked to an entity summary table that provides more detailed reference information. (B) Processing an HTML version of the article using Reflect illustrates OnTheFly faithfully reproduces Reflect annotations. (C) A closer look at the same ‘Materials and Methods’ section in PDF format converted by OnTheFly. Note that Reflect and ProMiner (Figure 4) identified the same gene name mentions and tagged the same false positive match for SLC.

MGI software evaluation process for text mining applications. MGI applied the basic steps for managing any major software acquisition project to evaluate potential text mining applications for gene indexing.

Visualization of ProMiner HTML and PDF tagging. The ProMiner tagged-entity visualization feature uses color-coded highlights to identify the source dictionary (mouse or human) of a tagged biological entity. We annotated this Journal of Immunology article with both the PDF and HTML versions of ProMiner to compare tagging styles and performance, then did a manual search for mouse, (highlighted in light blue) in Adobe Acrobat Reader and the HTML browser. As an example of a false positive hit, SLC is tagged as a synonym for the human gene CCL21; the actual reference is to Japan SLC, a mouse strain resources database. In MGI, this article was indexed to mouse genes Tlr4 and Ly96. (A) ProMiner 7.1 for PDF uses layers in Adobe Acrobat to flag gene names in the context of the original article layout. This makes it easier for MGI curators to scan specific sections of articles, such as Materials and Methods section and figure legends, for gene mentions. Due to issues related to PDF text extraction and conversion, this version of ProMiner has difficulty identifying some hyphenated terms and Greek symbols (such as the α in TNF-α), which are correctly tagged in the more mature HTML version of ProMiner. We provide feedback to SCAI on false negatives and false positives in specific documents, so they can enhance ProMiner processing rules and PDF labeling. (B) ProMiner 6.4 for HTML tags gene mentions using hypertext links and numerical references. Here, tan hypertext links indicate gene names matched to human dictionary terms, blue hypertext links are matched to mouse dictionary terms, those with two color tags are found in both dictionaries. Underlying hypertext links point to an object view window that displays the term reference ID and lists all gene synonyms. (C) In ProMiner 7.1 for PDF, human gene dictionary matches are labeled in red, mouse terms are in green, terms found in both dictionaries are highlighted in yellow or orange (there is no meaning associated with these different shades of highlighting; this is a labeling issue that will be addressed in a future update). Link-outs, outlined in red, indicate a popup window containing detailed information about the gene entity, synonyms, and source dictionary, is available by clicking on the term. A link-out for the tagged gene entity TLR4, identified in human and mouse source dictionaries from Entrez Gene, SwissProt and MGI, is shown.