Designing a tighter searching space for pairwise global sequence alignments over multiple scoring systems

The need to repeat alignments on a pair of amino acid residue sequences to select an appropriate scoring function and detect a significance of score arises in genomics and proteomics. While computing the alignments obtained through a set of typical scoring matrices with the corresponding default gap cost, we observe that many aligned segments are shared with a reference global alignment. We show that parameters extracted from the search for an alignment corresponding to a good scoring system can be used to predict the deviation of alignments computed with respect to different scoring matrices. By training sample pairs of protein sequences from SCOP 1.71 of the ASTRAL database, we build the approximated probability distribution of distance from a node on a reference path to the alignments based on other scoring schemes with respect to the proposed parameters. We show that the overall computational cost to perform alignments using `three' scoring matrices and the proposed method can be reduced to 11% of a normal Needleman-Wunch global alignment with an average 92% accuracy.