Neural and Statistical Classification to Families of Bio-sequences

In this paper we present a novel technique to compute feature vectors for use with artificial neural networks and other pattern recognition techniques that is designed for classifying families of biological sequences. Such sequences present unique challenges due to the fact that they vary in length and often consist of many symbols relative to the number of exemplars available. The latter property presents a specific challenge with respect to avoiding over generalization. We explore a novel approach involving computing the entropy of pair-wise correlations between co-occurring symbols in the strings to generate feature vectors which are of fixed size, much smaller than the original string lengths, and still effective at discerning differences between classes of strings. We apply the technique and show its effectiveness on an RNA family classification problem.