Generating Features Using Burrows Wheeler Transformation For Biological Sequence Classification

Recent advancements in biological sciences have resulted in the availability of large amounts of sequence data (both DNA and protein sequences). The annotation of biological sequence data can be approached using machine learning techniques. Such techniques require that the input data is represented as a vector of features. In the absence of biologically known features, a common approach is to generate k-mers using a sliding window. A larger k value usually results in better features; however, the number of k-mer features is exponential in k, and many of the k-mers are not informative. Feature selection techniques can be used to identify the most informative features, but are computationally expensive when used over the set of all k-mers, especially over the space of variable length k-mers (which presumably capture better the information in the data). Instead of working with all k-mers, we propose to generate features using an approach based on Burrows Wheeler Transformation (BWT). Our approach generates variable length k-mers that represent a small subset of k-mers. Experimental results on both DNA (alternative splicing prediction) and protein (protein localization) sequences show that the BWT features combined with feature selection, result in models which are better than models learned directly from k-mers. This shows that the BWT-based approach to feature generation can be used to obtain informative variable length features for DNA and protein prediction problems.