The alignment bootstrap method tests the statistical significance of a pair alignment by simulation. Sequence-randomized samples of each of the two sequences are repeatedly derived and subjected to a full pair alignment using parameters identical to the real alignment. The alignment scores from these randomized alignments are collected and drawn as a histogram. These alignment scores are local (Smith-Waterman) scores and they use the specified score matrix, so they may not be identical to the parent alignment score. The alignment score you are testing, computed with identical parameters to the boostrap tests, is shown above the bootstrap histogram.
The random sampling of the sequences for each test alignment is done in a manner that retains the exact length and amino acid composition of the parent sequences, so that only the residue order is randomized. Testing indicates that the scores from these randomized alignments conform approximately to a Gaussian distribution with a small positive skew, though the mean and standard deviation vary of course.
I will try in the future to add a simple statistical interpretation of the result. For now, you should be able to rely on eyeball analysis of the bootstrap histogram as real alignments generally are well off to the right (highly signicificant).
James H. Thomas, Department of Genome Sciences, University of Washington