Clarification on high Nfail counts in modkit pileup output despite valid coverage

[dasidhartha2018]

Hi modkit team,

I am analyzing CpG methylation using modkit pileup on Oxford Nanopore data
(mapped BAMs, reference-guided).

Across multiple samples, I observe that the majority of reads contributing to
coverage end up in the Nfail column, even when valid_coverage ≥ 3.

Some details:

• Genome: Plasmodium falciparum 3D7

• Command used (example):

  modkit pileup
  --reference PlasmoDB-64_Pfalciparum3D7_Genome.fasta
  --modified-bases C
  --combine-mods
  input.bam output.bed

• In the resulting bedMethyl files:

  • valid_coverage is often high (≥3 for many CpGs)
  • count_modified and count_canonical are low
  • most reads appear to be classified as Nfail
  • fraction of methylated CpGs among covered sites is very low

This behavior is consistent across samples and across different coverage cutoffs.

My questions:

1. Is a high Nfail count expected behavior in cases of low-confidence or low-level CpG methylation?
2. Which filters most commonly cause reads to be classified as Nfail?
   (e.g. modification probability threshold, basecall quality, alignment flags, context mismatch)
3. Is there a recommended way to summarize or interpret datasets where
   valid coverage exists but most reads fail confidence filters?
4. Would adjusting parameters like min-mod-prob be appropriate to explore this further?

The behavior seems biologically plausible for this organism, but I would like
to confirm that my interpretation of Nfail is correct.

Thanks for the great tool, and for any clarification!

[ArtRand]

Hello @dasidhartha2018,

Is a high Nfail count expected behavior in cases of low-confidence or low-level CpG methylation?

No. The estimated threshold should discard about 10% of the total calls. What is the estimated threshold value? You can see it in the console log or using --log $log_file.

You may also try running with --no-filtering, although this isn't the recommended approach, it might help figure out what's going on. You may see high numbers of N_fail (lower confidence modification calls) if the modifications of proximal to other modifications.

Which filters most commonly cause reads to be classified as Nfail? (e.g. modification probability threshold, basecall quality, alignment flags, context mismatch)

The Nfail counts are per-site-per-read when the call probability is below the threshold value. Documentation can be found here

Is there a recommended way to summarize or interpret datasets where valid coverage exists but most reads fail confidence filters?

As mentioned above, you may try running with --no-filtering.

Would adjusting parameters like min-mod-prob be appropriate to explore this further?

I would try running $ modkit modbam sample-probs --hist -o $out and look at the distributions. If you post them here I can take a look as well.