Thread by @rrwick

I have finally updated my @nanopore basecalling comparison with the current tools! Instead of living on GitHub, it's now a proper preprint:
biorxiv.org/content/10.110…

Some highlights follow...
(1/12)

Guppy is my overall favourite. It has good accuracy and lots of features, and (if you can run it on a GPU) is blazingly fast.
(2/12)

Guppy's new flip-flop model is slower but more accurate. It especially does well with read-level accuracy.
(3/12)

Modified bases in native DNA can be a big source of consensus errors. Our main test set was K. pneumo which has Dcm methylation at the CCAGG/CCTGG motif, and most of the errors in our assemblies were in that motif.
(4/12)

Basecalling with a custom model trained (using Sloika) on other K. pneumo solved this problem! Dcm-related errors dropped to near zero.
(5/12)

Generalised conclusion: if you can train a custom model using reads from your species of interest, your basecaller will be able to handle the DNA modification in that species, giving better accuracy (especially in the consensus).
(6/12)

This applies to sequencing native DNA. Amplified DNA should be modification free and probably does just fine with the default basecalling models.
(7/12)

We also trained a model using a bigger neural network than Guppy's default model. It did a lot better in both read and consensus accuracy, but it was slow.
(8/12)

Generalised conclusion: Guppy's neural network architecture is making a compromise between speed and accuracy. Higher accuracy is possible if you're willing to accept slower basecalling.
(9/12)

If you're using Guppy and working with native DNA from Enterobacteriaceae, you should download our custom models and give them a try:
monash.figshare.com/articles/Train…
(10/12)

Training custom models with Sloika wasn't easy: lots of time and effort to get the data ready, filter reads, set thresholds, control RAM usage, etc. It worked, but many @nanopore customers won't have the time or resources to do it.
(11/12)

So finally, a request of @nanopore: make Sloika easier to run, please! 😀 If you do, more users can train and share custom Guppy models for different organisms.
(12/12)