Literature-based discovery: Finding implicit associations between genes and diseases.

摘要

Swanson has argued that due to explosive growth in publications and overspecialization in the sciences certain types of knowledge based on known logical premises that share common concepts remain undiscovered. Establishing linkages between such logical premises by applying automated or semi-automated techniques to identify evidence in published literature is called literature-based association discovery. Although the potential of literature-based association discovery has been widely recognized, several key problems remain. First, the approaches are largely exploratory and rely on heuristics. Second, the approaches make use of annotations of keywords or mere word (co-)occurrences in text, considering word variations and synonyms at most. Third, they under-utilize or ignore available resources, such as online ontologies or thesauri. Lastly, the approaches lack strong validity as they have not been extensively evaluated. This study attempts to address these key problems. Targeting gene-disease associations as a testbed, we propose a modeling approach based on the Bayesian inference network whereby genes and diseases are represented as nodes and are connected via two types of intermediate nodes, namely gene functions and phenotypes. To estimate the probabilities involved in the model, two learning frameworks are compared. The first framework uses an established set of gene-disease associations. The second framework employs a larger but weaker set of gene-phenotype associations. Also, two learning techniques, one baseline scheme that uses co-annotations of keywords and the other that takes advantage of online free-text information, are presented. Our proposed approach is evaluated on a benchmark data set created from real-world data. The evaluation demonstrated that (a) the proposed approach is effective for predicting specific gene-disease associations, (b) use of free-text information consistently produces better performance than the use of keyword annotations, (c) information acquired from full-text documents can significantly improve discovery as compared to information extracted only from abstracts, and (d) domain ontologies can be leveraged to enhance discovery of associations between gene functions and phenotypes.