Mining sequence features for DNA-binding site prediction

Protein-DNA interactions play pivotal roles in gene regulation and DNA replication and repair. Since the 3-dimensional structure of most proteins is still unknown, computational methods which can identify DNA-binding sites from protein sequences are in demand. In this study, we used a greedy method to search for features that are useful for the identification of DNA-binding sites. 5 features were selected from a pool of 534 features. Using the selected 5 features, a Naive Bayes method achieved 0.31 Matthews correlation coefficient (MCC), which is an improvement over a previous method that used only 2 features as input. Since all of the 5 features can be derived from protein sequences, the proposed method can identify DNA-binding sites using only protein sequences as input.