Protein structure prediction is a challenging task and many different methods have been proposed to address it. More specifically, various computational techniques can be used to provide relations of newly discovered proteins to other proteins with known properties. By determining how amino acid sequences are related to those of known proteins,we can recognize their fold catergory and thus make predictions for their structure and function . HMMs are commonly used in fold recognition and also demontsrate high performance.

Protein fold recognition using a reduced state-space hidden Markov model
A hidden Markov model (HMM) with a reduced number of states, which simultaneously learns amino acid sequence and secondary structure for proteins of known three-dimensional structure, is used for structure prediction. A HMM with a reduced state–space topology is introduced. The model employs an efficient architecture with a small number of states and a low complexity training algorithm. Secondary structure information is introduced to the model to increase its performance and it is used in such a way that allows the use of the low complexity algorithm. The number of states is equal to the number of the different possible formations of secondary structure. The model is trained using the low complexity likelihood maximization algorithm for each candidate fold. The obtained results when multi-class fold recognition is performed with this model are equivalent or even better than other similar approaches. The major advantage of the proposed approach is that the computational load of the model is significantly smaller than conventional methods based on full HMM. In this case only the primary sequences of proteins are needed in the test set.

Certain improvements are also introduced to that model that further ameliorate its fold recognition performance without increasing its complexity. These improvements take place in two steps. More specifically, in the first step the number of states is decreased by adopting the simple {E,H,L} alphabet for the secondary structure. This reduction leads to an even smaller number of parameters that need to be calculated in the training phase and simultaneously to better results. In the second step, the predicted or the true secondary structure sequences is additionally used in scoring the test set sequences. Thus the use of the complex forward algorithm for scoring is avoided and also the exploitation of the secondary structure information of the test set proteins is attainable. The predicted secondary structure sequences are calculated with the use of PSIPRED. The obtained results show that when the number of states is reduced, the fold recognition accuracy is improved. Moreover, when the predicted secondary structure sequences are employed the fold recognition accuracy further increases and it becomes even better, apparently, when the true secondary structure sequences are used, as in that case there are no errors in the secondary structure.