Speeding up minimap2

Question

[colindaven]

Dear all,

has anyone done any benchmarking on speeding up long read alignment algorithms ?

I mainly use minimap2, but its' runtime varies by a factor of 10 across our cluster. I've been trying mm2-fast https://github.com/bwa-mem2/mm2-fast, the partially accelerated version, but without much success so far.

Is for example PAF output faster than SAM ?

Have others worked out how to scale minimap2 for Promethion scale datasets ? I expect LRA https://github.com/ChaissonLab/LRA has a similar runtime from their presented results, and others seem slower still (ngmlr etc).

Thanks

[GenoMax]

You are already using multiple threads and are asking for an additional speedup? You could split the data files up and start multiple jobs in parallel as a sledgehammer solution. I have never worked with Promethion size data so don't have a direct insight there.

[colindaven]

The input splitting is an interesting idea, which comes at the expense of using far more CPU resources. I'll give this a try but also benchmark with some other options.

edit: I am using 24 threads as I have found that to be fastest on my infrastructure when using hyperfine for benchmarking.

Thanks

Answer 1

[shelkmike]

1) Yes, when Minimap2 makes paf files it is faster than when it makes sam files.
2) For a speedup at cost of accuracy you can increase the minimizer length ("-k") and window length ("-w").
3) You can increase "-I". If the reference is larger than 4 Gbp, this will accelerate Minimap2 at cost of increased RAM consumption.

Also, see https://github.com/lh3/minimap2/issues/322

[Buffo]

For a speedup at cost of accuracy you can increase the minimizer length ("-k") and window length ("-w").

Exactly, I wouldn't do that if the sequencing reads come from PacBio (CLR) or oxford nanopore.

[shelkmike]

By default, for Nanopore reads Minimap2 uses -k 15 -w 10. This combination of parameters was intended for old Nanopore reads which often had an average accuracy below 90%. I think that for modern Nanopore reads, which have an average accuracy >95%, a user can increase -k and -w to some extent without compromising the alignment accuracy.

[Buffo]

Yes, I know. My point is that the quality of the data not only depends on the sequencing chemistry, it also includes the sequencing performance, the library preparation, the input DNA, etc. I would increase the -k for HiFi reads and/or when the coverage is high.