Using Background Knowledge to Improve Inductive Learning of DNA Sequences

Successful inductive learning requires that training data be expressed in a form where underlying regularities can be recognized by the learning system. Unfortunately , many applications of inductive learning| especially in the domain of molecular biology|have assumed that data are provided in a form already suitable for learning, whether or not such an assumption is actually justiied. This paper describes the use of background knowledge of molecular biology to re-express data into a form more appropriate for learning. Our results show dramatic improvements in classiication accuracy for two very diierent classes of DNA sequences using traditional \oo-the-shelf" decision-tree and neural-network inductive-learning methods. The goal of inductive learning is to take a collection of labeled \training" data and form a classiier that accurately predicts the labels of future data. To be successful it is essential that the data be expressed in a form where the learner can recognize underlying regularities. This problem of nding an appropriate encoding for training data is widely recognized as being central for all inductive learning eeorts. However, most work in inductive learning has traditionally assumed that the data are initially provided in a form that is already appropriate for learning (whether or not such an assumption is actually justiied). In the notable exceptions when such an assumption is not made, the most common approach has been to create a set of features carefully engineered for the particular learning problem. For example, Quinlan 19] documents his two-person-month eeort deening an appropriate set of features for learning the chess We thank Steve Norton and Lorien Pratt for helpful discussions and comments. end-game of lost 3-ply with a decision-tree learner. Similarly, Kamm and Singhal 10] describe the process by which diierent methods for sampling a speech signal were explored to nd the one that yields the best encoding of data for learning by a neural-network learner. The representation of training data is particularly important when learning classiiers for DNA sequences. In this domain training data are usually represented as strings over a four-letter alphabet; each character represents a nucleotide, which is the fundamental unit of the DNA polymer. All past learning eeorts have attempted to learn directly from these raw data. However , this representation only expresses low-level information about nucleic-acid sequences and does not encode any important features that biologists nd useful in describing regions of DNA. We began with the assumption that the use of an appropriate …