<p>For the vast majority of proteins no experimental information about the three-dimensional structure is known, but only its sequence. Therefore, the easiest way to obtain some understanding of the structure and function of these proteins is by relating them to well studied proteins. This can be done by searching for homologous proteins. It is easy to identify a homologous sequence if the sequence identity is above 30%. However, if the sequence identity drops below 30% then more sophisticated methods have to be used. These methods often use evolutionary information about the sequences, which makes it possible to identify homologous sequences with a low sequence identity.</p><p>In order to build a three--dimensional model from the sequence based on a protein structure the two sequences have to be aligned. Here the aligned residues serve as a first approximation of the structure.</p><p>This thesis focuses on the development of fold recognition and alignment methods based on evolutionary information. The use of evolutionary information for both query and target proteins was shown to improve both recognition and alignments. In a benchmark of profile--profile methods it was shown that the probabilistic methods were best, although the difference between several of the methods was quite small once optimal gap-penalties were used. An artificial neural network based alignment method ProfNet was shown to be at least as good as the best profile--profile method, and by adding information from a self-organising map and predicted secondary structure we were able to further improve ProfNet.</p>

For the vast majority of proteins no experimental information about the three-dimensional structure is known, but only its sequence. Therefore, the easiest way to obtain some understanding of the structure and function of these proteins is by relating them to well studied proteins. This can be done by searching for homologous proteins. It is easy to identify a homologous sequence if the sequence identity is above 30%. However, if the sequence identity drops below 30% then more sophisticated methods have to be used. These methods often use evolutionary information about the sequences, which makes it possible to identify homologous sequences with a low sequence identity. In order to build a three--dimensional model from the sequence based on a protein structure the two sequences have to be aligned. Here the aligned residues serve as a first approximation of the structure. This thesis focuses on the development of fold recognition and alignment methods based on evolutionary information. The use of evolutionary information for both query and target proteins was shown to improve both recognition and alignments. In a benchmark of profile--profile methods it was shown that the probabilistic methods were best, although the difference between several of the methods was quite small once optimal gap-penalties were used. An artificial neural network based alignment method ProfNet was shown to be at least as good as the best profile--profile method, and by adding information from a self-organising map and predicted secondary structure we were able to further improve ProfNet.

Indiana University-Purdue University Indianapolis (IUPUI) / Computational protein design aims at designing amino acid sequences that can fold into a target structure and perform a desired function. Many computational design methods have been developed and their applications have been successful during past two decades. However, the success rate of protein design remains too low to be of a useful tool by biochemists whom are not an expert of computational biology. In this dissertation, we first developed novel computational assessment techniques to assess several state-of-the-art computational techniques. We found that significant progresses were made in several important measures by two new scoring functions from RosettaDesign and from OSCAR-design, respectively. We also developed the first machine-learning technique called SPIN that predicts a sequence profile compatible to a given structure with a novel nonlocal energy-based feature. The accuracy of predicted sequences is comparable to RosettaDesign in term of sequence identity to wild type sequences. In the last two application chapters, we have designed self-inhibitory peptides of Escherichia coli methionine aminopeptidase (EcMetAP) and de novo designed barstar. Several peptides were confirmed inhibition of EcMetAP at the micromole-range 50% inhibitory concentration. Meanwhile, the assessment of designed barstar sequences indicates the improvement of OSCAR-design over RosettaDesign.