Rasmus Kirkegaard 20 May, 2021
We were lucky to be offered to try the new “Q20” chemistry ahead of the wide release. This post shows what the data looked like at the time we got it and might not reflect the performance when the kit enters a wide release as basecalling algorithms etc might have been fine tuned in the meantime. The main selling point with this chemistry+bonito basecalling was improved single pass accuracy. To evaluate the accuracy we sequenced the zymo mock community where a known reference genome is available for all of the organisms so it is possible to determine how well the basecalled reads reflect the genomes. The sequencing was carried out on the Promethion P24 platform with the Zymo mock DNA directly from supplier without any pretreatment (no short read elimination). The DNA fragments have a peak around 7 kbp but with a tail of short fragments. To maximise yield the short fragments would ideally have been removed but since the DNA material was limited I omitted this as I figured that a Promethion Flowcell would anyway generate plenty of data to check the quality of the reads for this low complexity sample.
Yes the overall modal alignment accuracy was already 99 % with this release when tested with the zymo mock community. There are still some species specific variations but the read quality is generally really good. The best chromosome contig recovered had an indel rate of 0.74/100kbp and mismatch rate of 0.37/100kbp, which is pretty mind blowing for a nanopore only assembly!!! The estimated ANI value was 99.9989 %! Polishing this metagenome assembly with racon did not seem to improve the quality of the best contig. A medaka model was not available and will likely be tricky to develop for contigs with this quality directly from the assembler. I assume that improvements in the basecalling models will bring the others on par with this in the future.
I have uploaded the basecalled data and the fast5 files to the ENA PRJEB43406. For illumina data I grabbed the reads from Lomans mock community blogpost.
The DNA size distribution was checked using a genomic screentape.
The reads were basecalled using bonito (v. 0.3.5) with a model specifically trained for this new type of data.
The modal alignment accuracy (reported by bonito) seems to be better than 99 % (Q20) from reading single DNA molecules!!! So the chemistry clearly delivers excellent performance on accuracy.
Since the zymo includes a number of different organisms it would be interesting to see if the performance was equally good for all organisms. The alignment accuracy distributions seems indicate that the basecalling is slightly better for some organisms than others. For many of the organisms the peak of the distribution is slightly higher than 99 % but for e.g. Salmonella Enterica the performance seems to be a bit lower. The models for basecalling are likely going to be improved before a wide release but it is important to keep in mind that your favourite pet organisms might have modifications that are not yet included in the basecalling models and thus show lower performance.
While it is great to see improvements of the single read accuracy. It is also interesting to see how the consensus accuracy is improving. To do this I ran and assembly with metaflye (v. 2.8.3-b1695) inputting the reads as –nano-corr and three of the chromosomes assembled into nice circular contigs even with the “short” DNA fragments from the zymo library. Here the assembly graph is visualised using bandage. I used the three chromosomes for comparison
IDEEL analyses (https://github.com/mw55309/ideel)
To ensure that indel errors left from the nanopore data do not cause genes to break we can compare the gene length with that of the best hit to a reference gene. Most genes appear to have a length that is similar to the best “BLAST” (DIAMOND) hit in the reference database indicated by a relative gene length of 1.
