Predicting protein structure with Bayesian statistics and models of evolutionary information.

Knowledge of the structure of a protein is essential for understanding the protein's function, but experimental structure determination requires large amounts of time and resources. Obtaining amino acid sequence information, however, is relatively easy, and there is substantial experimental evidence showing that this sequence information is sufficient for specifying the structure for many proteins. This situation motivates the development of theoretical and computational techniques for predicting protein structures from amino acid sequences. We have developed a probabilistic formalism for predicting local features of protein structure such as solvent accessibility and secondary structure. Our method uses Bayes' theorem and a consideration of the structural basis of inter-residue correlations. In addition it employs alternative descriptions of local protein structure by including combined categories of secondary structure and solvent accessibility along with structural information from nearest neighbors in the sequence. These techniques yielded 3-state secondary structure and 2-state solvent accessibility accuracies of 67% and 71%, respectively. This performance is competitive with the best previous methods based on datasets of non-homologous single sequences. We also developed two methods for using multiple sequence alignments to improve prediction accuracies. Both models were optimized for structure prediction by maximizing the mutual information between the model parameters and a database of local structure information. The first model consisted of a set of discrete substitution classes to which multiple sequence alignment sites could be assigned. This approach considered only the range of allowed amino acid substitutions at each alignment site. Use of this model attained 2-state solvent accessibility accuracies as high as 75%. A second probabilistic model was developed which took into account the number of each type of substitution occurring at a particular location. Use of the probabilistic model in secondary structure predictions achieved accuracies around 72%. This is equivalent to the best previous methods based on multiple sequence alignments. Our method is simpler, more cohesive, and less computationally costly than previous methods. It is also more physicochemically intelligible. All of the parameters in the prediction formalism and the evolutionary models are open to inspection and analysis regarding their implications about sequence-to-structure relationships in proteins.Ph.D.Applied SciencesBiochemistryBiological SciencesBiophysicsComputer sciencePure SciencesUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/130852/2/9811207.pd