The Manipulation of Medical Literature as Interference of Medical Sciences in Order to Formulate New Hypotheses

In this study we present a theoretical framework that can be used to explore the possibilities of automatic hypothesis generation from medical literature. We use the model of disconnected logic developed by D.R. Swanson. The assumption is that there exist complementary but disconnected literatures from which new information can be obtained when these literatures are connected. The theoretical framework provides a set of heuristics rules which are tested with empirical data derived from a medical literature database. The field of medical and pharmaceutical sciences can be divided into many subdisciplines which are more or less related amongst each other. The resulting scientific literature of these subdisciplines has strong intra-subdisciplinary connections, but weak inter-subdisciplinary connections. This means that knowledge from the one subdiscipline is not always integrated in the scientific research of another subdiscipline. We intend to interfere in these sciences by offering a theoretical model and the computational tools that provide a way to obtain new hypotheses from existing literature, by connecting knowledge from the different subdisciplines. Recent developments in computer text analysis have made it feasible to process large amounts of texts and thus address the question of automatic literature based scientific discovery in medical and pharmaceutical research. This study combines the novel approach of actively reconstructing implicit or hidden logical inference patterns in scientific literature with recent approaches to statistical and information-theoretical analysis of (large amounts of) raw medical texts. We explore and model the development of an automatic literature-based drug discovery system, which consists of three components: (1). The model of disconnected logic The model disconnected logic (Swanson, 1986; 1987; 1988; 1990) assumes that articles in medical and pharmaceutical literature are connected by logical relations. An actual case has been identified by Swanson (1986). He found that in one context of medical research it was known that dietary fish oil (= A) causes certain blood and vascular changes (=B). Whereas in another context scientific research it was recognized that patients suffering from Raynauds syndrome (=C) benefit from certain blood and vascular changes (i.e. those referred to under B!). However, the two areas of medical research proceeded without mutual recognition, and this phenomenon of two not-connected ’literatures’ has been called disconnected logic. The interesting case is that the connection was actively made by Swanson, and consequently it was tested and verified in further medical research. In principle, one could start the procedure with words that denote C, then start backtracking in medical literature databases by retrieving all words denoting B that co-occur with C-words, and then taking each B-word-candidate separately, retrieving all possible A-B word pairs. In practice, the methods used by Swanson were semi-automatic, that is, partly statistical and computational, partly by reading titles and abstracts of records from scientific literature. We believe that this approach can be modeled and simulated on well-known examples from scientific literature. In Vos (1991, 1992, 1993, 1994) and Kuipers, Sie and Vos (1992) discovery patterns in innovative pharmaceutical research have been analyzed. One inference type which we take as a prototype of how pharmaceutical innovation develops, is the transformation of an adverse drug reaction (ADR) into a new indication area for a particular drug. For example, the hypotensive effect of one drug as a side effect in the treatment of gastric ulcer patients might be profitable in the treatment of hypertension. This inference type is quite similar to the syllogistic construction described above. (2) The methodology of automatic wordcategorization The methodology consists of methods and techniques to determine possible candidates of the C-, Band A-components and their links (AB-, BC, and finally AC-connections) (cf. Rikken, Vos 1994). Our first experiment has been Adverse Drug Reactions (ADR) in medical research. ADRs can be defined unwanted, unexpected and undesired effects of drugs and the aim of the scientist is to minimalize the influence of ADRs o that the risk/benefit balance of a medical treatment remains on the positive side. We want to show that we can reformulate these ADRs by From: AAAI Technical Report SS-95-03. Compilation copyright © 1995, AAAI (www.aaai.org). All rights reserved. connecting them with knowledge from other subdisciplines into promising new hypotheses which can result in new knowledge. In the presented work the ADRs are studied in their potency of being candidate B’s as formulated in the Swanson model. (3) New methods in natural language processing Exploration of the Natural Language Environment (NLE) of candidate terms with the help of the CLARIT natural language processing-statistical software (Carnegie Mellon University (Prof.D.A. Evans) (Rikken, Evans, Vos 1994). We have made a computational linguistic translation of the simple syllogistic model of Swanson. This translation has resulted in a theoretical model of the use of NLEs in order to find and describe candidate B terms. We explore the NLE of concepts (e.g. ADRs) in different medical subdisciplines and combine these, in those eases where the NLEs have changed. The result is a new hypotheses, combining two statements with the ADR as an intermediate. An interesting case for modeling automated-based drug discovery is the recent development of diabetic nefropathy as a new indication for captopril. Captopril, was developed as the first ACE-inhibiting drug in the early 1980s for the treatment of hypertension and heart failure. Amongst the side-effects known for captopril was proteinuria. In the course of the 1980s it became clear that eaptopril was also beneficial in the treatment of proteinuria in diabetic nefropathy. Both the adverse drug reaction and the beneficial effect of captopril are widely recognized phenomena. Thus, we have a model case at hand to reconstruct the discovery pattern proposed in the project. In preliminary research we were able to find the relevant documents in the literature base before the researchers in the field actively made the connection. The technological status of the project is that we can automatically process large amounts of medical doeuments with both statistical and computational linguistic tools. However, further development of specific algorithms to reduce the number of terms of the thousands of potentially relevant word-combinations found at the several devision steps in the automatic text based discovery procedure, requires further creative computer research.