How to compare k-mer completeness between phased genomes

[desmodus1984]

Hi,
My group used PacBio HiFi reads to assemble a genome, and then we further scaffolded it with Hi-C data, and got the pseudo-haplotype and two phased genomes.
I first ran the k-mer QV analysis for the pseudo-haploid since it was the one that was released to us first, and I got yesterday the phased.
I read somewhere about shared/private k-mer plots, and I would appreciate if you could tell me how to do this.
I don't have parents, since we work with an elusive species, and finding parent-offspring pairs is very hard.

I am running the regular merqury analysis of the two genomes with code:

merqury.sh Hlep.meryl hap1.fasta hap2.fasta Phased-haps

and I wondered about this haplotype-level analysis, since it is mentioned right-away in the log:

Merqury -haps - Job started: Tue Sep 16 13:55:49 EDT 2025
read: Hlep.meryl

No haplotype dbs provided.
Running Merqury in non-trio mode...

asm1: hap1.fasta
asm2: hap2.fasta
out : Phased-haps

Any help and tips very appreciated.

I also wanted to inquire about one idea.
I am interested/concerned/curious about improvements of genome assemblies. As I mentioned, I QV'ed the scaffolded/non-scaffolded genomes, and I was surprised of the result:
final_assembly 5521 2454370272 69.7014 1.07117e-07
HLep.fx 5521 2454370312 69.7014 1.07117e-07
Both 11042 4908740584 69.7014 1.07117e-07
The QV and error rates of them are exactly the same. Isn't it too unexpected?
This surprises me because the scaffolded genome (final_assembly) is a little longer (2900 bp), has more contigs/less scaffolds, but more and bigger gaps. With these changes, I was expecting the scaffolded assembly to have lower QV but I was wrong.
Also, I wanted to inquire about the impact of read-error correction + genome polishing. Many studies mention genome polishing but not read polishing. For instance, I was thinking on using DeepConsensus (https://www.nature.com/articles/s41587-022-01435-7) to correct the reads, which should increase the number of solid kmers, and then polish the genome.

[arangrhie]

Hi @desmodus1984 ,

I read somewhere about shared/private k-mer plots

Do you mean the spectra-asm plots? That should've been generated automatically. Look for file names with spectra-asm.
The plot is showing what portion of the kmer in each assembled hap1.fasta and hap2.fasta are unique, shared, or missing. This could be used as a rough indicator of which haplotype is missing what, and if you see 'shared' portion in the expected 1-copy kmer peak, that means your assembly contains some false duplications.

and I wondered about this haplotype-level analysis, since it is mentioned right-away in the log:

The log is telling no parental dbs are provided, and thus running Merqury in non-trio mode.

The QV and error rates of them are exactly the same. Isn't it too unexpected?

If the only change between the unscaffolded and scaffolded is the introduction of gaps, then this is as expected. Merqury doesn't determine gaps as error bases. Gaps are gaps.

Using error corrected reads instead of raw-reads

Unfortunately, all error correction algorithms out there still has a chance of introducing errors in the corrected read, which will lead to a new false k-mer. For example, what's often detected in an error-corrected read is a 'smashed' version of the two haplotypes in a heterozygous region. I haven't done a close evaluation yet on the latest versions, since the error correction algorithms are evolving fast - might be worth evaluating at some point again. For now, I'd be a little more conservative and use the raw reads.

HiFi reads are an exception, they are generated with DeepConsensus behind the scene. However, these are generated per-SMRT bell which eliminates the chance of haplotype smashing. Still, it has its own biases, so my preference is using a hybrid of Illumina / Element + HiFi kmers. It slightly overestimates the QV, but so far is the closest to get as a proxy in a reasonable time for genomes which a benchmark genome doesn't exist.

Best,
Arang