Protein secondary structure prediction with semi Markov HMMs

Secondary structure prediction has been an essential task in determining the structure and function of proteins. Prediction accuracy is improving every year towards the estimated 88% theoretical limit (Rost. B., http://cubic.bioc.columbia.edu/papers/2002 rev dekker/paper.html). There are two approaches for the secondary structure prediction. The first one, ab initio (single sequence) prediction, does not use any homology information. The evolutionary information, if available, is used by the second approach to improve the prediction accuracy by a few percentages (Schmidler, S.C. et al., J. Computational Biology, vol.7, no.1/2, p.233-48, 2000). We address the problem of single sequence prediction by developing a semi Markov HMM, similar to the one proposed by Schmidler et al. We introduce a better dependency model by considering the statistically significant amino acid correlation patterns at segment borders. Also, we propose an internal dependency model considering right to left dependencies without modifying the left to right HMM topology. In addition, we propose an iterative training method to estimate the HMM parameters better. Putting all these together, we obtained 1.5% improvement in three-state-per-residue accuracy.