How to improve the QV of the hifiasm assembly?

[zmz1988]

Dear prof. Heng Li,

Thank you so much for the tool! I have been using hifiasm since 2021, and am a big fan of it. I recently run into different results from different versions of hifiasm, and would like to ask your opinion in solving the problem, if you don't mind. So my situation is as below:

Recently I assembled all of my old data with the new hifiasm v0.19.6. The data includes hifi (~25 x) plus Nanopore (50k or 30k length, >50x). Then I compared the resulted assemblies with my old assemblies assembled by hifiasm v0.16.5 (hifi reads only).

I found that QVs (k-mer based QV calculated by Merqury, k-mer data base calculated from hifi reads) are much lower in the new assemblies (e.g. 60 in the old vs. 45 in the new), despite the superb NG50 values of the new assemblies. The completeness (k-mer based calculated by Merqury) is also relatively low (e.g. 98% in the old vs. 96% in the new, some new ones even have completeness 75%).

I'm not sure why this happen, as the data for both trials are the same. I mean some parts of the new assemblies are from Nanopore reads, which may cause inconsistency with the hifi reads and lower the QV. But I don't know how to explain the lower completeness (especially 70% ones) in the new assemblies.

I would prefer to use the newly assemblies by hifiasm v0.19.6 for our publication, because they are gapless from telomere to telomere. So my question is: whether we have a way to improve the QV? Would you recommend two runs of racon with hifi reads to polish the assemblies, though I know it's recommended not to polish the final assemblies from hifiasm?

I'm very sorry for writing such a long post. Thank you very much in advance!

Some more info of the assemblies: Species: Arabidopsis thaliana Code: hifiasm -o ${sample} -t 10 --ul ${sample}_ont_50k_reads.fq.gz --ul-rate 0.2 --primary ${hifi_fastq_file} &> ${sample}_hifiasm.log