Learning to predict DNA hydration patterns

Abstract

The authors examine the problem of learning to predict hydration patterns around DNA molecules. It is assumed that there is a limited, but so far unknown, set of hydration patterns, and that there is a set of features of a DNA molecule which determines its pattern. Since the patterns for the DNA molecules in the database were not known a priori, most traditional classifier learners cannot be applied directly. The authors have combined cluster analysis with a decision tree learner to develop classifiers, even though training examples were not initially labeled with classes. Some empirical results of this learning are presented, and it is shown how the learned decision trees are being used to gain insight into the domain of DNA crystallography.<>