The success of protein structure modeling by homology requires an accurate sequence alignment between the query sequence and its structural template. However, sequence alignment methods based on dynamic programming (DP) are typically unable to generate accurate alignments for remote sequence homologs, thus limiting the applicability of modeling methods. A central problem is that the alignment that would produce the best structural model is generally not optimal, in the sense of having the highest DP score. Suboptimal alignment methods can be used to generate alternative alignments, but encounter difficulties given the enormous number of alignments that need to be considered. We present here a new suboptimal alignment method that relies heavily on the structure of the template. By initially aligning the query sequence to individual fragments in secondary structure elements (SSEs) and combining high-scoring fragments that pass basic tests for 'modelability', we can generate accurate alignments within a set of limited size.
Chapter 1 introduces the field of protein structure prediction in general and the technique of homology modeling in particular. One subproblem of homology modeling—the sequence to structure alignment of proteins — is discussed in Chapter 2. Particular attention is given to descriptions of the size, density and redundancy of alignment space as well as an explanation of the dynamic programming technique and its strengths and weaknesses. The rationale for developing alternative alignment techniques and the unique difficulties of these methods are also discussed.
Chapter 3 explains the methodologies of S4 — the alternative alignment program we developed that is the main focus of this thesis. The process of finding alternative alignments with S4 involves several steps, but can be roughly divided into two main parts. First, the program looks for combinations of high-similarity fragments that pass basic rules for modelability. These 'fragment alignments' define regions of alignment space that can be searched more thoroughly with a statistical potential for a single representative for that region. The ensemble of alignments that is thus created needs to be evaluated for accuracy against the correct alignment. Current methods for doing so, as well as adjustments to those methods to better suit the realm of remote homology alignments, are discussed in Chapter 4. A novel measure for determining similarity between alignments, termed the inter-alignment distance (IAD) also is developed. This measure can be used to assess quality, but is also well-suited to finding redundant alignments within an ensemble.
In Chapter 5, the results of testing S4 on a large set of targets from previous CASP experiments are analyzed. Comparisons to the optimal alignment as well as two standard alternative alignment methods, all of which use the same similarity score as S4, demonstrate that S4's improvement in accuracy is due to better sampling and filtering rather than more sophisticated scoring. Models made from S4 alignments are also shown to significantly improve upon those made from optimal alignments, especially for remote homologs. Finally, an example of a sequence to structure alignment offers an in depth explanation of how S4 finds correct alignments where the other methods do not.
Chapter 6 describes a set of three experiments that paired S4 with the model evaluation tool ProsaII in a homology modeling pipeline. There were two primary objectives in this project. First, we wanted to test different methods for finding remote homologs that could serve as input to S4. And second, we evaluated the use of ProsaII as a method for discriminating between good and bad models, and thus also between homologous and non-homologous templates. The first two experiments are essentially blind searches for homologous sequences and structures. The third experiment takes remote templates returned by PSI-BLAST and uses S4 and ProsaII to find alignments and determine whether the template is a structural homolog. While S4 was able to find homologs in the blind searches, the alignment/model quality and level of discrimination was found to be higher when the input to the pipeline came from a set of structures produced by a template selection method.
Finally, Chapter 7 discusses the consequences of this research and suggests future directions for its application.