String Searching Methods for Bioinformatics
Kalle Karhu · Aaltodoc (Aalto University) · 2013
The cost of obtaining biologically relevant data via sequencing has been declining rapidly, far surpassing the decline in computing costs. This is highlighting a need for more efficient, and thus cheaper, ways to analyze all of this data. Analyzing such data commonly requires searching through the text representing it in one way or another. The focus of this thesis is on improving the efficiency of the computational approaches that one may wish to use when searching through such texts.More precisely, it addresses three subproblems related to text searches in bioinformatics. First, we consider the approximate, indexed alignment of long sequences. We present an approach using an index that combines q-sampling and block addressing for the initial approximate location of promising alignments, which are then studied more carefully using a multi-pattern, q-gram algorithm. Based on our experimental results, this approach is able to answer alignment queries notably faster than previous approaches, using only a fraction of the memory required by them. We additionally show that the quality of alignments and even the exon mappings produced by this approach are not worse than those produced using previous approaches. Second, we consider indexed multi-pattern matching. For this subproblem, a set of multiple patterns is preprocessed, speeding up our search of this set from an index structure. This thesis presents the first experimental results on this type of an indexed, multi-pattern matching setting together with new theoretical insights. Practical approaches to this setting are presented, and our experimental results suggest that the presented approaches to preprocessing notably improve later searches from the corresponding index structures. Namely, compressed suffix arrays and bidirectional FM-indexes are considered in our study. Finally, we consider protein motif discovery. We present a new graph-theoretical approach based on de Bruijn graphs. Moreover, we show how to further improve the query times of this approach using similarity indexing. Our experiments suggest that the presented approaches produce motif predictions of equal quality notably faster than previous methods.