A modified stack decoder for protein secondary structure prediction

Secondary structure prediction is an important step in determining the structure and function of proteins. A fundamental assumption of current Bayesian secondary structure prediction methods is the conditional independence of residues which occur in distinct segments. This assumption enables the exact calculation of posterior probabilities by using pre-determined probabilistic models. However, this assumption is clearly violated in the case of protein sequences due to the existence of structural motifs which rely on sequentially distant segments interacting in three-dimensional space, including β-sheets. It has been suggested that the inability to capture such nonlocal interactions may be the main reason for the low accuracy typically achieved in β-strand prediction (Schmidler, S.C. et al., 2000; Frishman, D. and Argos, P., 1996). Furthermore, current Bayesian segmentations are based on maximum a posteriori or marginal posterior mode searches, which return a single segmentation that is optimal in some sense. We introduce a new secondary structure prediction method based on a modified version of the well-known stack decoder. The proposed method is an N-best search algorithm, which enables us to use the returned multiple segmentations to improve over a single segmentation. Also, due to the way the segmentations are constructed, it is possible to exploit the non-local interactions between β-strands in a sub-optimal way with the ultimate goal of increasing the overall prediction accuracy.