Inter-species normalization of gene mentions with GNAT

Abstract

The test data set, lexica, and links toexternal data are available at http://cbioc.eas.asu.edu/gnat/

Efficient Extraction of Protein-Protein Interactions from Full-Text Articles

“…Thus, we backed off from some approaches presented previously, such as GNAT for entity mention normalization and an alignment-based pattern matching algorithm (see [20], [19]). The focus of all building blocks in our systems is on annotation time, and our goal was to provide a service that can handle a full-text article in 10 seconds and still have reasonable accuracy; our most accurate composition/tuning parameters should still be able to analyze a full text in about two minutes.…”

“…NER systems in recent years have tended towards primarily employing machine learning techniques, including conditional random fields, due to the consistently high performance these techniques provide when trained on a high-quality corpora such as the BioCreative II gene mention corpus [22]. In contrast, however, systems for identifying proteins (also called normalization and grounding, EMN) have largely focused on dictionary-based techniques; some notable recent systems include GNAT [20] and GeNo [23]. For BioCreative II.5, we have identified several attributes which are useful for supporting the identification of proteins found, including: 1) the ability to associate a confidence with each mention found; 2) improving consistency via enforcing a one-senseper-document assumption; 3) generating a list of candidate proteins (identifications) to which each mention could refer.…”

“…We then narrow down each list of IDs per protein mention by species. We have shown previously [19], [20] how genes/proteins can be disambiguated by using context profiles, that is, information available on each protein that can be compared to the current text and measure the overlap in GO terms, disease associations, tissue specificity, chromosomal location of genes, protein length and mass, and so on. For efficiency required in the online tasks in BioCreative II.5, however, we do not perform actual disambiguation, but rank IDs by dictionary match and species only.…”

“…For both INT and IPT this can simply be achieved by submitting multiple identifiers per protein, thus leading to hundreds of UniProt IDs per article. The highest AUC iP/R submission of 43.5% contained an average of 83 IDs for each of the 252 relevant proteins (20,888 IDs in total for the 61 relevant documents, 342 IDs on average per document), with a resulting precision of 1.2%. For a database curation scenario, narrowing down the number of IDs per protein/article might prove suitable, for curators as well as authors writing an SDA; thus, we focus on reporting F-scores in the remainder.…”

“…For related work, we refer the reader to the other articles published in this special issue, together with the overview paper by Hirschman et al [15]. Some of the approaches we discuss build on earlier systems described in [19] (gene mention normalization; pattern generation) and [20] (inter-species gene mention normalization). We discuss related work inline with the methods, data sets, and discussion where appropriate.…”