Documentation request for `dmr` technique and console messages

[SuhasSrinivasan]

Hi @ArtRand

Happy holidays!
Thank you so much for developing and maintaing modkit!

Here is an excerpt from a dmr run for DRS.

> reading reference FASTA at "/mnt/drive1/references/ont/gencode.v48.transcripts.fa"
> 55992 common sequence(s) between FASTA and both samples
> running single-site analysis
> using default prior, Beta(α: 0.55, β: 0.55)
> estimating max coverages from data
> sampled 4396558 a records and 4475828 b records, calculating max coverages for 95th percentile
> calculated max coverage for a: 569 and b: 646
> calculated max coverage 569 is greater than maximum allowed (100), setting to 100
> calculated max coverage 646 is greater than maximum allowed (100), setting to 100
> errors:
+--------------------------+---------+
| error                    | count   |
+--------------------------+---------+
| missing-in-one-condition | 5175396 |
+--------------------------+---------+

> finished, processed 31823062 sites successfully, 5175396 failed

Query 1
It would be very helpful to get descriptive information about the steps being performed and the console messages shown.

Query 2
The above excerpt is from comparing two replicates of the same cancer cell line.
a. Is it typical to have millions of modifications different between replicates?
b. I have observed missing-in-one-condition in the millions range while comparing two cancer cell lines, is that typical?

Query 3
Is the differential modification analysis normalized to coverage?
If so, can you please provide additional information about the approach?

[ArtRand]

Hello @SuhasSrinivasan,

It would be very helpful to get descriptive information about the steps being performed and the console messages shown.

Sure. I agree having interpretable logs is important. Have you looked at the log file (--log $file_path), it has more detail.

The above excerpt is from comparing two replicates of the same cancer cell line.
a. Is it typical to have millions of modifications different between replicates?

Could you qualify what you mean by "different"? I have some code that can generate correlations between runs, maybe this would be helpful. One thing I've been meaning to do is make a little simulation that defines a probability of methylation at each CpG in the genome then samples calls at each position. (I appreciate you're using transcriptome here, but the principal is the same). My intuition is that even when the underlying probability of methylation is the same, you will observe some positions that end up "different" just by chance. If you're trying to get a measure of reproducibility, what we generally do is look at the Pearson's R and RMSE between the %-methylation of two replicates requiring a minimum valid coverage at each position in the intersection.

I have observed missing-in-one-condition in the millions range while comparing two cancer cell lines, is that typical?

I can look at some of our internal data to see how typical this is.

Query 3
Is the differential modification analysis normalized to coverage?

It isn't "normalized", but at lower coverage the model attempts to capture uncertainty in the data. It looks like your run has very high counts (>500), at that number of observations I would expect the empirical %-modification to be pretty close to the frequency in the sample.

If so, can you please provide additional information about the approach?

There are details on the methods in the documentation

Please let me know if you have any further questions or if you would like that code that performs the correlation analysis.

[SuhasSrinivasan]

Hi @ArtRand

Happy New Year!

Thank you for reviewing and for the additional details.

Could you qualify what you mean by "different"?

I want to understand what is the sensitivity and error rate of (1) Nanopore DRS, (2) dorado and (3) modkit dmr , to identify millions of RNA modifications to be different between any two replicates of the same cell line?
I am new to this and it seemed high for inter-replicate, while inter-cell line makes biological sense.

I reviewed the EPI2ME RNA modifications best practices for information about accuracy and modkit validate, but it does not go into depth about differential modifications.

I can look at some of our internal data to see how typical this is.

That would be very helpful, looking forward to it.

At that number of observations I would expect the empirical %-modification to be pretty close to the frequency in the sample.

Thanks, then it is ok to report findings from modkit dmr as is ?