Predicting Secondary Structure
hidden Markov models
Hidden Markov models representing protein secondary structure not only provide a method of predicting secondary structure classification from primary structure, they also have few enough parameters that they can be interpretable by inspection. While determining secondary structure requires only deciding between the three categories helix, sheet, and loop (sometimes slightly more specific categories are used) HMMs recognize much more complexity than this. For example, a structure prediction for a single amino acid might describe a helix location that is oriented towards the interior of the protein. Or a prediction might specify a loop structure that is very close to a helix. HMMs are naturally depicted by directed, weighted graphs and the analysis of features such as cycles, vertex degree, and weights all can be related to secondary structure.<\p>
Our HMM predictor is based on the development and analysis of a 36-state HMM described by Martin, Gibrat, and Rodolphe in Choosing the optimal hidden Markov model for secondary-structure prediction.
Our work is in secondary_predictor_martin.py in
the prediction directory.