Training Set Design for Pattern Discovery With Applications to Protein Motif Detection

Abstract

Supervised pattern discovery techniques have been successfully used for motif detection. However, this requires the use of an efficient training set. Even in cases where a lot of examples are known, using all the available examples can introduce bias during the training process. In practice, this is done with the help of domain experts. Whenever such expertise is not available, training sets are usually picked at random. We present a new strategy for designing good training sets that uses phylogenetic trees to automatically reduce the bias in training sets. When applied to helix-turn-helix motif detection, we show that the technique improved the error rates over a random choice of training set. More specifically, false positive rates show a marked improvement.