Automated Theory Formation in Bioinformatics

A theory learned by an inductive logic programming (ILP) system such as Progol [5] usually comprises a set of concepts, expressed as logic programs, which can be employed for a classification task. This classifying ability can, in turn, be used for prediction tasks. A scientific theory, however, comprises much more information: concepts; hypotheses relating concepts; explanations and empirical justifications of the hypotheses; representation schemes; experimental methodologies and so on. Working mainly in mathematics, we have used the HR system [1] to form theories about some objects of interest in a domain. For example, in group theory, where the objects are groups, HR invents concepts, makes conjectures about those concepts, and proves (some of) the conjectures using the Otter theorem prover [4]. Despite it’s history in mathematics, we have developed HR as a domain-independent machine learning program. In particular, the format for background information is very similar to that for Progol. Given this, we are currently exploring various possibilities for automated theory formation (ATF) using bioinformatics datasets. We describe here an application of HR to the mutagenesis data set [6] and suggest some advantages of ATF over ILP, some disadvantages, and some possibilities for the fruitful combination of the two techniques.