When a chain is split during base-level alignment, its chaining score is
reduced. However, the chaining score of its suboptimal chain remains the same.
This leads to underestimated mapping quality.
[PMID:18688272] shows that the base following GT tends to be A or G (i.e. R) in
both human and yeast, and that the base preceeding AG tends to be C or T (i.e.
Y). In the new model, we pay no cost to GTr..yAG, but we pay half of the cost
if there is no r or y. This improves the junction accuracy when mapping to
human and mouse and decreases the accuacy when mapping to SIRV. My guess is
that SIRV does not honor this trend. Need to investigate in future.
Also in this commit, --cost-non-gt-ag is aliased to -C. The default is changed
to 9 instead of 5. I also added --splice-flank to enable the above model. This
may become the default once I confirm my hypothesis on SIRV.
though it might not work in an extremely rare case: the end of a sequence ends
at X*16384 and it is the last sequence in a batch. This can be resolved by
never letting the kstream_t buffer empty.
In the default splice mode, minimap2 applies two rounds of spliced alignment:
first assuming GT-AG to be the splice signal across all splicing sites and then
assuming CT-AC to be the signal. This is the idea strategy.
In the MM_F_SPLICE_BOTH mode, minimap2 applies one round of spliced alignment,
assuming GT-AG and CT-AC to be the splice signals AT THE SAME TIME. This will
be faster but less accurate. I don't think anyone would like to run minimap2 in
this mode, so I am removing it for clarity.
It turns out the key problem is not the minimizer density. It is the max
occurrence that tends to affect results more, especially sensitivity. There is
still lots of work to do, but for now, it seems a good start.
Heng Li
mvdbeek