Email updates

Keep up to date with the latest news and content from BMC Bioinformatics and BioMed Central.

Open Access Software

Meta-Alignment with Crumble and Prune: Partitioning very large alignment problems for performance and parallelization

Krishna M Roskin1*, Benedict Paten2 and David Haussler3

Author Affiliations

1 Department of Computer Science, Univ. of California, Santa Cruz, USA

2 Center for Biomolecular Science & Engineering, Univ. of California, Santa Cruz, USA

3 Howard Hughes Medical Institute, Univ. of California, Santa Cruz, USA

For all author emails, please log on.

BMC Bioinformatics 2011, 12:144  doi:10.1186/1471-2105-12-144

Published: 10 May 2011

Abstract

Background

Continuing research into the global multiple sequence alignment problem has resulted in more sophisticated and principled alignment methods. Unfortunately these new algorithms often require large amounts of time and memory to run, making it nearly impossible to run these algorithms on large datasets. As a solution, we present two general methods, Crumble and Prune, for breaking a phylogenetic alignment problem into smaller, more tractable sub-problems. We call Crumble and Prune meta-alignment methods because they use existing alignment algorithms and can be used with many current alignment programs. Crumble breaks long alignment problems into shorter sub-problems. Prune divides the phylogenetic tree into a collection of smaller trees to reduce the number of sequences in each alignment problem. These methods are orthogonal: they can be applied together to provide better scaling in terms of sequence length and in sequence depth. Both methods partition the problem such that many of the sub-problems can be solved independently. The results are then combined to form a solution to the full alignment problem.

Results

Crumble and Prune each provide a significant performance improvement with little loss of accuracy. In some cases, a gain in accuracy was observed. Crumble and Prune were tested on real and simulated data. Furthermore, we have implemented a system called Job-tree that allows hierarchical sub-problems to be solved in parallel on a compute cluster, significantly shortening the run-time.

Conclusions

These methods enabled us to solve gigabase alignment problems. These methods could enable a new generation of biologically realistic alignment algorithms to be applied to real world, large scale alignment problems.