Classification of Protein Sequences Using Markov Models Classification of Protein Sequences Using Markov Models

Preface This study was completed as a master thesis at the institute of Informatics and Mathematical Modeling (IMM)-Technical University of Denmark (DTU). The project has been completed in the period between September 1 st 2003 and Marts 1 st 2004 (and is rated as 35 points). Associate Professor Paul Fischer also IMM, DTU has been supervisor for the project. The report deals with a general classification problem in the area of bioinformatics, namely secondary structure prediction. However since we do not have a biological background, the problem is treated as a mathematical classification problem. The report presents the results of the conducted analyses, conclusions and implementations. The report describes the analyses, methods and ideas to a level of details, making it possible to reconstruct the tests and the results presented in this text. The structure of the report is chronological and may be read from one end to the other. However some of the more theoretical sections may be closely connected to the corresponding result sections. Several different notations are used in the report. Literature references are given as a number in index parentheses [] in a smaller font-size, like [1]. Footnotes appear in superscript and usually in the end of a sentence, like this 1. Equations, formulas and other expressions are numbered using a notation like (X.Y which is the Yth expression in section X. Abstract This project deals with a specific classification problem in the area of bioinformatics and biology. The problem, typically referred to as secondary structure prediction deals with how the structure of protein sequences may be classified using a number of predefined structure classes. This project analyses the possible use of Markov models for this classification problem. Markov models are statistical models which may be used to infer the different structure classes for protein sequences based on some training data. The performance of the developed models are compared to other known models in the area, specifically the GOR models, which are similar to Markov models since they are both statistical models. The obtained results show that Markov models may be used for secondary structure prediction achieving better performances than just guessing at the most frequent structure class. Starting out with a simple Markov model able to predict around 51% of the structures correctly, the model has been extended and combined with other methods resulting in a prediction accuracy of 57.2% (an increase of around 6%). …