Improved BLAST

Altschul et al. suggested in 1997 [2] an improved BLAST algorithm that allows indels in the alignment.

The algorithm stages follows:

1.

When considering the dynamic programming matrix to align two strings, we search along each diagonal for two w-length words such that the distance between them is $\leq A$ and their score is $\geq T$. Whenever we encounter such a pair of hits we concatenate them.

2.

In the second stage we want to allow local alignments with indels, similarly to the FASTA algorithm. We allow two local alignments from different diagonals to merge into a new local alignment composed of the first local alignment followed by some indels and then the second local alignment. This local alignment is essentially a path in the dynamic programming matrix, composed of two diagonal sections and a path connecting them which may contain gaps. Unlike in FASTA, where we only allowed the diagonal to have local shifts, restricted to a band, here we allow local alignments from different diagonals to merge as long as the resulting alignment has a score above some threshold. This method results in an alignment structure which is much less regular.