Exploring the Role of Knowledge Representation and Reasoning in Biomedical Text Understanding

There is considerable effort being devoted to mining information from medical and scientific literature, in particular, from Medline abstracts and from full-text articles. Such information is being used, for example, to reconstruct biological pathways, identify pathogenic mechanisms and, importantly, to identify functional relationships that can be used to predict disease onset and its course thereafter. Our interest is in exploring the role of knowledge representation and reasoning (KR&R) as it relates to the problem of understanding biomedical text. The role we envision for a KR&R system in this context is as a knowledge store for a small, focused subset of abstracts gleaned from the Medline corpus that is relevant to a problem of interest. The system will infer new knowledge from the represented abstracts which can then be stored in a larger data repository and reported to a biologist. We are specifically interested in designing a system that, given a set of abstracts, can perform many of the same inferences that a biology expert would make if given the same set of abstracts. Inferences that go beyond the predictions of the biologist are particularly interesting, but our initial goal is to emulate the biologist. We have selected the disease neurofibromatosis type 1 (NF1) for our study with the goal of developing a model that can be applied to reasoning about other diseases and problems of interest. This approach is focused narrowly on a particular problem and as such may not lead to solutions relevant for general problem solving. However, we believe there is an important role for specialized problem solvers in the larger context of biomedical text understanding. Working closely with a domain expert in biology is providing valuable insights into how the computational synthesis of information might best serve the needs of a biologist. This preliminary report describes our current work on hand analysis and translation of abstracts and our proposed overall approach to the problem.