R9.4 2D error rate

[thermophile]

I'm interested in starting to use the minion for meso-throughput targeted sequencing (1.5-3kb amplicons). I haven't heard error rates for the R9 since just after they were released (Nick Loman's blog). What are people getting in the wild? If you can fully overlap the 2d reads, what's the consensus sequence error rate?

[Ola]

I'm interested in starting to use the minion for meso-throughput targeted sequencing (1.5-3kb amplicons). I haven't heard error rates for the R9 since just after they were released (Nick Loman's blog). What are people getting in the wild? If you can fully overlap the 2d reads, what's the consensus sequence error rate?

Good question. For 1D it is ~10-12% now, somewhat sample/sequence and alignment dependent. That would give you ~5-fold more data compared to 2D runs, but if you plan for barcoding and don't need the troughput 2D gives slightly lower error rates (can be ~3% but more often 6-10).

Consensus errors (from many reads) are mainly around short homopolymers but other sequences can show up as heterozygotes as well before polishing.

[thermophile]

Thanks. I'm afraid that's not good enough yet for what I want to do. But that's good to know.

[Ola]

Thanks. I'm afraid that's not good enough yet for what I want to do. But that's good to know.

I checked error rates (bwa mem -x ont2d + picard CollectAlignmentSummary) for an amplicon run now, with 145 2D reads for a 8 kb human region. It gave 1.5% subst error and 3.2 % indel errors. This was reads from end of run, and with high coverage we can quality-filter reads if a lower average error rate is needed. Taking only the top 20% of reads (base qv > 25 on average) yields error rates of 0.66% for subst and 2.3 % for indels.

[gringer]

What do you want to do?

[thermophile]

taxonomic surveys-I generally want to look at many samples shallowly. I'm considering doing 16/18S + ITS1 to allow for the error rate but would need to do more bioinformatic pipeline writing than I currently have ability/time. I want to do this but think I'll push it to back burner for a few months till either the error rate drops or I have time/find a student to work on some processing pipeline that could leverage both ITS and 16s for clustering.

[Brian Bushnell]

PacBio CCS reads work great for full-length 16S sequencing. Typically ~99.5% accuracy for the best read in each cluster.

[thermophile]

Thanks Brian, I didn't mention that I'm trying to come up with a way to get longer reads that we can do inhouse and we don't have pacbio. I'm still waffling. It's cheap enough to try that I may give it a shot and try tempering my expectations.

[gringer]

For taxonomic surveys, you'd be better off doing a full metagenomic analysis instead of looking at specific genes. The MinION is more than adequate for this purpose.

Metrichor provides a WIMP ("What's In My Pot") workflow for doing exactly this. It uses Kraken* behind the scenes to vote on the origin of subsequences of each read, generating a consensus taxon for each read. The end result is a tree with counts that will show in real time (i.e. during sequencing) the content of your sample. The output using extracted, unamplified DNA correlates extremely well with the expected output for "canned" metagenomic samples:

https://nanoporetech.com/publications/metrichor-applications-whats-my-pot-wimp

* it used to use Kraken, but that may have changed with recent releases

[wdecoster]

The error rate is also species-dependent. If I'm not mistaken the basecaller is mainly trained on the e.colli and lambda genome. Improvements to base calling for human genomes will happen if training is also applied to those genomes. (I believe this is being worked on).