Improvement in Fold Recognition Accuracy of a Reduced-State-Space Hidden Markov Model by using Secondary Structure Information in Scoring

Fold recognition is a challenging field strongly related with function determination which is of high interest for the biologists and the pharmaceutical industry. Hidden Markov Models (HMMs) have been largely applied for this purpose. In this work, the fold recognition accuracy of a recently introduced Hidden Markov Model with a reduced state-space topology is improved. This model employs an efficient architecture and a low complexity training algorithm based on likelihood maximization. Currently we further improve the fold recognition accuracy of the proposed model in two steps. In the first step we adopt a smaller model architecture based on {E,H,L} alphabet instead of DSSP secondary structure alphabet. In the second step we additionally use the predicted and the correct secondary structure information in scoring of the test set sequences. The dataset, used for the evaluation of the proposed methodology, comes from the SCOP and PDB databases. The results show that the fold recognition performance increases significantly in both steps.