Finding functional sequence elements by multiple local alignment.

Algorithms that detect and align locally similar regions of biological sequences have the potential to discover a wide variety of functional motifs. Two theoretical contributions to this classic but unsolved problem are presented here: a method to determine the width of the aligned motif automatically; and a technique for calculating the statistical significance of alignments, i.e. an assessment of whether the alignments are stronger than those that would be expected to occur by chance among random, unrelated sequences. Upon exploring variants of the standard Gibbs sampling technique to optimize the alignment, we discovered that simulated annealing approaches perform more efficiently. Finally, we conduct failure tests by applying the algorithm to increasingly difficult test cases, and analyze the manner of and reasons for eventual failure. Detection of transcription factor-binding motifs is limited by the motifs' intrinsic subtlety rather than by inadequacy of the alignment optimization procedure.