Mining MEDLINE for Similar Genes and Similar Drugs

Hypothesis generation, a crucial initial step for making scientific discoveries, relies on prior knowledge, experience and intuition. Chance connections made between seemingly unrelated concepts sometimes turn out to be fruitful. A key goal in text mining is to assist in this process by automatically discovering a small set of interesting hypotheses from a suitable text collection. We focus on text mining in the biomedical domain using MEDLINE, the database produced by the National Library of Medicine with more than 12 million citations. Our overall goal is to build applications that mine MEDLINE for novel concept connections and thereby support scientists in hypothesis discovery. In this paper we first present concept profiles as a mechanism for generating concept representations from text collections. There are several advantages offered by concept profiles. They can be as current as the text database or they can be generated from temporal subsets. Profiles may be restricted to particular views and also they may be generated for concepts that are as complex as needed. We then show how concept profiles may be used to identify similar concepts. In particular, we present experiments where concept profiles are used to identify genes that are associated with the same disease and drugs that are functionally similar.