Segment-based multiple sequence alignment

In this PhD thesis the segment-based approach for multiple sequence alignment, initially introduced by the DIALIGN program, is thorougly investigated and substiantially improved. The segment-based approach belongs to the class of local alignment methods and thus is very strong in finding locally conserved motifs, whereas global methods align the input sequences globally from the beginning to end without specifically looking at locally occuring conserved motifs. Local alignments and especially segment-based methods therefore play an important role in molecular biology research, which is underscored by the fact that the results of this PhD thesis have already been extensively used in various biological research areas. Initially we present a complete re-implementation of the DIALIGN approach in chapter 3 – DIALIGN-T – which also embraces several improvements, such as the exclusion of low-scoring sub-fragments and weight score factors yielding a statistical superior method on local and global benchmark databases. However DIALIGN-T still uses a greedy and, therefore, very naive strategy to build the final alignment so that in chapter 4 we re-formulate the assembling phase as an optimization problem that is NP-complete, but for which we can proove it to be a fixed parameter tractable (FPT) in the number of input sequences, under reasonable assumptions. Since we are interested in approaches that are useful in practice, we develop a plane-sweep algorithm that optimally solves the assembling problem whereby its computational time basically grows with the number of simultaneously occuring conflicting situations. By exploiting the ideas of the plane-sweep algorithm, we extend it, in chapter 5, to a full algorithmic framework which acts as a basis for developing further optimal or near-optimal heuristics for assembling an alignment from a given set of input similarities. Inspired by this framework, we improve, in chapter 6, DIALIGN-T to its most recent version DIALIGN-TX, which incorporates substiantial improvements by combining greedy and progressive strategies for assembling the alignment. In order to measure the quality of our improvements, we used the standard benchmark databases BAliBASE and BRaliBASE II on global alignments and the artifically generated databases IRMBASE and DIRMBASE on local alignments. The results show that DIALIGN-TX is currently outperforming all other methods on the local benchmark databases while still providing very good results on global alignments, i.e. it even outperforms the very popular global alignment program CLUSTAL W on the global benchmark database BAliBASE 3. Altogether we conclude that DIALIGN-TX is one of the strongest methods on the important class of local alignments while still providing very good results on global alignments and consuming in practice only a reasonable amount of computational time. In combination with the algorithmic framework we obtain a rich basis or future improvements to the segment-based approach for computing general and (biological) domain-specific multiple sequence alignments.