Biomedical Association Discovery via Complementary TDM

Ever-growing textual data make it increasingly difficult to effectively utilize all the information relevant to our interests. For example, Medline—a comprehensive bibliographic database in life science—currently contains over 16 million articles, and the number is rising rapidly by 2,000–4,000 per day. Given the substantial volume of information, it is crucial to develop intelligent information processing techniques, namely, text data mining (TDM), that could help us manage the information overload. While its importance is clear, the definition of TDM has been somewhat vague in use within the related research communities. In a broader sense, it refers to every text processing equipped with some intelligence, such as information extraction, text categorization, and summarization, whereas in a more restricted sense of TDM, it refers only to the discovery of heretofore unknown knowledge [1]. The most significant difference between the two is that the former focuses on the information explicitly stated in text, and the latter on implicit knowledge which could be discovered by synthesizing fragments of information extracted from a large volume of textual data. Historically, these two types of TDM have been tackled in rather separated research communities. In our ongoing research projects, we are in part developing a TDM (in the more restricted sense) system to discover unknown gene-disease associations through an analysis of the biomedical literature [2]. The system is based on an extension of an information retrieval model and is capable of predicting causative genes for a given heredity disease. Independently, we are also working on automatic annotation of gene functions represented by Gene Ontology (GO) terms based on textual description extracted from journal articles [4], which was partly motivated by the TREC Genomics Track 2004 and the first BioCreative workshop. The annotation framework consists of flexible gene name matching and text categorization modules and is currently regarded as the state-of-the-art. This latter work can be considered the broader sense of TDM, attempting to abstract semantic knowledge embedded in natural language text. Although the aforementioned projects, gene-disease association discovery and automatic GO annotation, are characterized by different aspects of TDM, they are closely related and may work complementarily since genes are the critical entities in both frameworks. The next sections describe some details of these two projects and discuss how they could be integrated to help each other.