DIALIGN-T: An improved algorithm for segment-based multiple sequence alignment

Background: We present a complete re-implementation of the segment-based approach to multiple protein alignment that contains a number of improvements compared to the previous version 2.2 of DIALIGN. This previous version is superior to Needleman-Wunsch-based multi-alignment programs on locally related sequence sets. However, it is often outperformed by these methods on data sets with global but weak similarity at the primary-sequence level. Results: In the present paper, we discuss strengths and weaknesses of DIALIGN in view of the underlying objective function. Based on these results, we propose several heuristics to improve the segment-based alignment approach. For pairwise alignment, we implemented a fragment-chaining algorithm that favours chains of low-scoring local alignments over isolated high-scoring fragments. For multiple alignment, we use an improved greedy procedure that is less sensitive to spurious local sequence similarities. To evaluate our method on globally related protein families, we used the well-known database BAliBASE. For benchmarking tests on locally related sequences, we created a new reference database called IRMBASE which consists of simulated conserved motifs implanted into non-related random sequences. Conclusion: On BAliBASE, our new program performs significantly better than the previous version of DIALIGN and is comparable to the standard global aligner CLUSTAL W, though it is outperformed by some newly developed programs that focus on global alignment. On the locally related test sets in IRMBASE, our method outperforms all other programs that we evaluated.