Read quality

[marchoeppner]

After getting everything working (thanks again for the quick code fix), I noticed that for my test data the read quality that is being modelled is pretty terrible ;)

We get a reasonable high fraction of Q20 bases these days; whereas squigulator - even on the r10 DNA model - is much more pessimistic (< 20% Q20). Is that something that I can influence with parameters somehow? wasn't quite clear from the documentation.

[hasindu2008]

I guess if you try the sup basecalling, accuracy may improve. Also, --dwell-std 0 to squigulator can improve accuracy a bit more.

But still not as much as you would like to, because the pore model used for stimulation for r10.4.1 is not the greatest quality. Was planning to work on this, but never got time to work on it.

May I know what is your use case here?

[marchoeppner]

Sure. We are developing a sequencing-technology-agnostic pipeline for metabarcoding analysis. https://github.com/bio-raum/FooDMe2

However, it's been a challenge to develop a reasonable benchmark dataset for profiling pipeline performance across primer systems specifically for ONT. These barcodes are between 100 and 550bp in length (mostly 300-400). I understand that the shortest amplicons aren't currently supported by Squigulator, which is fine for the moment. But as mentioned above, we rely on the read data having somewhat reasonable quality scores as we do filter on that early in the processing chain. And right now the quality of the simulated data looks a bit rubbish ;)

[hasindu2008]

Hi @marchoeppner

When we developed Squigulator and did the benchmarks for the publication, the ONT flowcell chemistry was R9.4.1. For R9.4.1 chemistry, the squigulator has better similarity to the real data compared to other simulators that generated the raw signal, as we showed here https://genome.cshlp.org/content/34/5/778.full?sid=cd2c8aec-be46-4c9e-885c-8452ac069f64.

Then, ONT released this R10.4.1 chemistry 4KHz for which they provided a pore model with only the "mean values". We sequenced some modification-free data and obtained the standard deviation values, and that is what we have in squigulator right now. Then ONT released R10.4.1 5KHz, for which they did not release an updated pore model, so we are still relying on the 4KHz model. So the simulated signals are far from reality. For your case, do you require signal data or do you only need read data?

[marchoeppner]

Hi,

I can happily live without the basecalling part, sure. I suppose I could write some crude approximation of ONT behavior to simulate reads that way. But that sounds like work.. ;)

If it helps to understand the issue better, here are Quality plots from nanoplot for real data and the data simlulated with squigulator (R10, HAC 5.2.0)

Real:

Suigulator:

[hasindu2008]

Yes, the identity scores that I get for basecalled simulated reads for R10 are far from what is from real data.

For your use case, if the reads are sufficient, I believe you could also look into any read-level simulators for R10 (if any). For R10, I believe simulating at the read level is perhaps easier than doing so via signal level due to the lack of a good pore model at the moment, which was not the case for R9.