Minimap2 presets are perhaps too strict?

[schorlton]

Thanks for the great tool!

My understanding from the source is that the minimap2 presets are set based on --min-identity to the following:

quast/quast_libs/ca_utils/align_contigs.py

Lines 66 to 71 in 06874b8

	if qconfig.min_IDY < 90:
	preset = 'asm20'
	elif qconfig.min_IDY < 95:
	preset = 'asm10'
	else:
	preset = 'asm5'

However, I'm wondering if these are actually too strict as the --min-identity is described?

@lh3 would be able to comment best, but my understanding from the minimap2 manual and issues is asm5 can align up to a maximum of 5% sequence identity, but in fact, most of the sequence should be far less than this. Same thing for asm10 and asm20. Therefore, for average identity across an entire contig of 95%, asm5 is not the optimal setting and probably too strict.

I'd recommend just using asm20 for all values of --min-identity (except maybe 1% or less, where asm5 may be safe). The speed/memory trade-offs for alignment just aren't there if you're doing bacterial genome alignment, and you're post-filtering the alignments anyway.

[alexeigurevich]

Hi! Thanks for your highly valuable comment. Yes, it seems that we interpreted presets recommendations incorrectly. Let's wait for Heng response for a while (I am sure he is quite busy), but we should probably update our preset setting strategy.

[alexeigurevich]

In the commit above, I changed the default preset from asm5 to asm10 for the cases when minimum identity is >=95 and <99 (the default value is 95, can be changed via --min-identity). For values between 90 and 95, asm10 was already used, for value below 90, we always use asm20.

At the same time note, that the difference between presets is in just 4 variables: -w, -B, -O, -E (according to the minimap2 manual).

# asm5:  -w19 -B19 -O39,81 -E3,1   (Only use this preset if the average divergence is far below 5%)
# asm10: -w19 -B9  -O16,41 -E2,1
# asm20: -w10 -B4   -O6,26 -E2,1

But QUAST also uses its own -B and -O values on top of the presets (unless "--large" a.k.a. QUAST-LG is used):

quast/quast_libs/ca_utils/align_contigs.py

Lines 82 to 85 in 1c9db32

	additional_options = ['-B5', '-O4,16', '--no-long-join', '-r', str(qconfig.local_misassembly_min_length),
	'-N', num_alignments, '-s', str(qconfig.min_alignment), '-z', '200']
	cmdline = [minimap_fpath(), '-c', '-x', preset] + (additional_options if not qconfig.large_genome else []) + \
	['--mask-level', mask_level, '--min-occ', '200', '-g', '2500', '--score-N', '2', '--cs', '-t', str(max_threads), ref_fpath, contigs_fpath]

So the expected difference after my commits is not as dramatic. And this also means that our settings were not as strict as it seemed :)

[schorlton]

Thanks for your detailed response! Out of curiosity, why does QUAST use it's own -B and -O? Would it not be good to go with the ones built into minimap2 for assemblies?

[alexeigurevich]

There are several reasons:

1. Legacy: We added minimap2 in QUAST v.5 and before that we used the Nucmer aligner (from v.1.0 till v.4.6.3). To make the transition smoother for the users, we tried to configure minimap2 in such a way that its results were more or less "compatible" with the ones in "old" QUAST versions.
2. Results robustness: We don't want to demonstrate "dramatic" changes in the results when user simply changes the --min-identity threshold e.g., from 95 to 94 or from 90 to 89. So, on the one hand we switch presets, on the other hand we smooth the switch by using the same -B and -O for all of them.
3. QUAST goals and minimap2 goals are a bit different: the aligner tries to find good (in part, continues) alignments, i.e. show similarities between the two sequences, while we are also trying to find discrepancies (==assembly errors) and we don't want them to be "hidden" inside the long alignments :)

At the same time, I think we will change this approach in the future and will simply use the built in presets to make things easier and more straightforward for the users (we are working on a brand new QUAST now!).
Anyway, thanks for initiating this discussion!

[schorlton]

Thanks for the explanation! That makes sense and looking forward to seeing the update 🎉