Avoid MAPQ=60 for single-end multimappers

[marcelm]

by ensuring that we compute at least two alignments even if the first that is found has edit distance 0.

Closes #293

Because I still need to test whether this actually fixes #293, I’m marking this as draft.

[ksahlin]

approved.

[marcelm]

(Note that I had a typo/logic error in the PR and pushed a fix.)

I tried to do some measurements.

It’s a bit hard to inspect multimappers only, so I looked at incorrectly mapped reads instead, which should all be multimappers: Assuming a couple of things about how the read mapper works, these should have at least two roughly equivalent mapping locations (from the read mapper’s point of view): The correct one and the one the read mapper actually outputs. So they should get MAPQ=0.

So issue #293 is essentially that we get many incorrectly mapped reads that are assigned MAPQ>0, which should actually have been assigned MAPQ=0.

I generated a test dataset with know mapping locations by picking every tenth 100-mer from the Drosophila reference (about 14.4 million reads). Then I could count incorrectly mapped reads that have MAPQ>0 and MAPQ=0, respectively. For good measure, I also counted the correctly mapped reads with MAPQ>0 and MAPQ=0. The results are as follows:

Read category	main (40748fe)	this PR (097807b)	BWA-MEM
Incorrect and MAPQ>0 (bad)	9.87%	0.54%	0%
Incorrect and MAPQ=0 (good)	0.03%	9.36%	9.94%
Correct and MAPQ>0	90.08%	87.20%	86.63%
Correct and MAPQ=0	0%	2.88%	3.41%

I think the relevant numbers are the two bold ones: The "bad" alignments are much closer to zero and the number of "good" ones has gone up. Also, all four percentages are now closer to BWA-MEM.

Note that the percentage of "correct and MAPQ>0" goes down (and "correct and MAPQ=0" up accordingly). I assume these are mostly multimappers that strobealign by chance placed into the "correct" position (the one from which the k-mer was sampled). They previously got MAPQ=60, but since they are multimappers, they should get MAPQ=0, which they now do.

[ksahlin]

great improvement, approved!

[marcelm]

I forgot to add a changelog entry in this PR and took the liberty of pushing that directly to main without a PR.

[marcelm]

I realized I hadn’t check the performance impact of this change, so here are some measurements on 100 bp reads.

• Drosophila: ~3% slower
• Maize: ~9% slower
• CHM13: ~5% slower
• Rye: ~5% slower

These measurements were quite noisy, especially for rye.

[ksahlin]

Ok, worth it.

[marcelm]

Agreed!