Are all global alignment algorithms and implementations correct?

Abstract

Abstract.Pairwise sequence alignment is perhaps the most fundamental bioinformatics operation. An optimal global alignment algorithm was described in 1970 by Needleman and Wunsch. In 1982 Gotoh presented an improved algorithm with lower time complexity. Gotohs algorithm is frequently cited (1447 citations, Google Scholar, May 2015), taught and, most importantly, used as well as implemented. While implementing the algorithm, we discovered two mathematical mistakes in Gotohs paper that induce sub-optimal sequence alignments. First, there are minor indexing mistakes in the dynamic programming algorithm which become apparent immediately when implementing the procedure. Hence, we report on these for the sake of completeness. Second, there is a more profound problem with the dynamic programming matrix initialization. This initialization issue can easily be missed and find its way into actual implementations. This error is also present in standard text books. Namely, the widely used books by Gusfield and Waterman. To obtain an initial estimate of the extent to which this error has been propagated, we scrutinized freely available undergraduate lecture slides. We found that 8 out of 31 lecture slides contained the mistake, while 16 out of 31 simply omit parts of the initialization, thus giving an incomplete description of the algorithm. Finally, by inspecting ten source codes and running respective tests, we found that five implementations were incorrect. Note that, not all bugs we identified are due to the mistake in Gotohs paper. Three implementations rely on additional constraints that limit generality. Thus, only two out of ten yield correct results. We show that the error introduced by Gotoh is straightforward to resolve and provide a correct open-source reference implementation. We do believe though, that raising the awareness about these errors is critical, since the impact of incorrect pairwise sequence alignments that typically represent one of the very first stages in any bioinformatics data analysis pipeline can have a detrimental impact on downstream analyses such as multiple sequence alignment, orthology assignment, phylogenetic analyses, divergence time estimates, etc.