Constructing Genome Scale Suffix Trees

Suffix trees have been the focus of significant research interest as they permit very efficient solutions to a range of string and sequence searching problems. Given a suffix tree that encodes a particular string, it is possible to solve problems such as searching for a specific pattern in time proportional to the length of the pattern rather than the length of the string. Suffix trees can also support inexact matching by dramatically improving the performance of dynamic programming. Therefore, suffix trees may enable a number of large scale bioinformatics problems to be solved more efficiently than previously thought. However, these benefits presume that a suffix tree of sufficient scale can be constructed. An inherent difficulty in suffix tree construction is that the tree construction requires a semi random walk over the tree as it is constructed. Therefore very large trees that will not fit in memory could take an unacceptably long time to construct due to excessive page faulting. In this paper we present a linear time construction algorithm that can construct suffix trees larger than memory using discrete sub-trees. The sub-trees can be constructed in parallel. The performance of the algorithm is evaluated using suffix trees constructed for chromosomes 1 and 12 of the human genome.