When backtracking, bwa-short does not keep the detailed alignment or the exact
start and end positions. To find the boundary and the CIGAR, the old code does
a global alignment with a small end-gap penalty. It then deals with a lot of
special cases to derive the right position and CIGAR, which are actually not
always right. It is a mess.
As the new ksw.{c,h} does not support a different end-gap penalty, the old
strategy does not work. But we get something better. The new code finds the
boundaries with ksw_extend(). It is cleaner and gives more accurate CIGAR in
most cases.
stdaln.{c,h} was written ten years ago. Its local and SW extension code are
actually buggy (though that rarely happens and usually does not affect the
results too much). ksw.{c,h} is more concise, potentially faster, less buggy,
and richer in features.
The changes after r317 aim to improve the performance and accuracy for very
long query alignment. The short-read alignment should not be affected. The
changes include:
1) Z-dropoff. This is a variant of blast's X-dropoff. I orginally thought this
heuristic only improves speed, but now I realize it also reduces poor
alignment with long good flanking alignments. The difference from blast's
X-dropoff is that Z-dropoff allows big gaps, but X-dropoff does not.
2) Band width doubling. When band width is too small, we will get a poor
alignment in the middle. Sometimes such alignments cannot be fully excluded
with Z-dropoff. Band width doubling is an alternative heuristic. It is based
on the observation that the existing of close-to-boundary high score
possibly implies inadequate band width. When we see such a signal, we double
the band width.
I really hate this: complex and twisted logic for a nasty scenario that almost
never happens to short reads - but it may become serious when the reference
genome consists of many contigs.
On toy examples, the code seems to work. Don't know if it really works...
Heng Li