How an alignment with multiple HSPs is evaluated in blastn, blastp?

Question

[sajal]

When blastn finishes searching, it reports many alignments (hits or query-sequence match), many of these alignments consist of many hsps. So, when it picks the top alignments, how does it evaluate the alignments's score?

I know how each HSP is scored and how the score is used to compute evalue. But, how these individual HSPs contribute to the alignment's score? If there is no idea of score for an alignment, then how does blast decide which alignments to keep and report?

I was going with the idea that every alignment is judged with it's highest scoring HSP (or lowest evalue HSP). But when I do split-database query, I found some alignment with a pretty high scoring HSP does not get picked by the search when run on the whole database. Is there any sum-statistics in play when evaluating alignments?

Answer 1

[Jean-Karim Heriche]

Yes, it's based on the sum statistics. The significance of an alignment is derived from the sum of the selected HSPs scores (see this Karlin & Altschul paper for the ungapped version, this was also shown to work for gapped alignments)

[sajal]

Jean-Karim: Thanks! I came across this paper a while ago and was trying to find this in the blast code with no success. So, release note of BLAST+ 2.2.29: January 3, 2014 says:

"Ungapped BLAST no longer uses sum statistics by default. Recover old behavior with -sum_statistics ˆag."

I am not sure if it's in use for gapped alignment though.