A HIERARCHICAL METHOD FOR PROTEIN SECONDARY STRUCTURE PREDICTION
Dr Yann Guermeur, LIP6, Universite Paris VI

Abstract: Protein secondary structure prediction is a problem of central importance in predictive structural Biology. It has been studied intensively for more than thirty years. However, the best models currently available to perform this task suffer from a fundamental drawback: their prediction is only based on the contents of a small segment of the sequence surrounding the residue the conformational state of which is to be determined. This procedure is suboptimal, since it is known that distant interactions play a part in the determination of the secondary structure.

We present a new method for protein secondary structure prediction, designed to overcome the restrictions induced by the local approach. It is based on a hierarchical architecture made up of three levels. At the first level, several classifiers implementing different statistical inductive principles compute initial class posterior probability estimates, for each residue. These estimates are combined at the second level by an ensemble method performing a multivariate linear regression. Its outputs are postprocessed with a dynamic programming algorithm. The underlying hidden Markov model has three states, one for each conformational state. The observations are the residues of the primary structure. The specificity of the HMM lies in the modeling of state durations. Furthermore, a variation of Viterbi's algorithm, called the "N-Best algorithm", is implemented. It can either be used to provide the biologist with a set of alternative predictions, making it easier for him to make his final prediction, based on his own expertise, or to incorporate automatically higher level knowledge.

Also this system can still be significantly improved, its recognition rate is already equivalent to those of the current best prediction methods, which establishes the relevance of the hierarchical approach for protein secondary structure prediction.

This seminar was held at the Department of Computer Science, Royal Holloway, University of London on 26 March 1999.

back