òThe logical foundations of goal-regression planning in autonomous agentsó, Artificial In.pdfVIP

Hidden Markov Model for protein secondary structure Juliette Martin, Jean-Francois Gibrat, and Francois Rodolphe Unit´e Math´ematique Informatique et G´enome, INRA, Domaine de Vilvert, 78350 Jouy-en-Josas Cedex, France (e-mail: [Juliette.Martin,Jean-Francois.Gibrat, Francois.Rodolphe]@jouy.inra.fr) Abstract. We address the problem of protein secondary structure prediction with Hidden Markov Models. A 21-state model is built using biological knowledge and statistical analysis of sequence motifs in regular secondary structures. Sequence family information is integrated via the combination of independent predictions of homologous sequences and a weighting scheme. Prediction accuracy with single sequences reaches 65.3% and raises to 72% of correct classification with profile information. Keywords: α-helix, β-sheet, prediction. 1 Introduction Proteins are the main actors of living cells. Many cellular constituents are made out of proteins. Almost all enzymes are proteins, cellular pumps and motors are made out of proteins. The function of a protein strongly depends of its 3D-structure. For in- stance, enzymes need to have a tight spatial complementarity with their substrates (reaction partners). Thus knowledge of a protein structure gives relevant clues to its function. Since genome sequencing started, the even widening gap between the number of protein sequences and protein structures available in databases en- hances the utility of structure prediction methods. Because of the structure- function relationship, structures are more conserved than sequences during evolution and therefore different sequences can have the same 3D structure. Structure prediction methods fall into two categories: • comparative modeling if a related structure is known and can be