Seeking some clarification of the m6A predicted results

[YCCHEN23]

Hello,

I'd like to express my appreciation for your impressive m6A detection tools. I've recently found numerous m6A sites from my libraries and observed strong concordance among the results from biological replicates by using m6anet. For this analysis, I used the Arabidopsis model with default parameters, focusing on the common sites of replicates without any filtering process, resulting in more than 200,000 sites.

However, I have a few queries:

Q1: The Pearson's correlation between the results of Human- and Arabidopsis-model was only 0.29 in my data. I am puzzled as to why the predicted results from the two models differ so greatly. Regardless of whether it's Human or Arabidopsis, ribonucleic acid remains the same, correct?

(I pooled all the four replicates, the throughput was more than 8 million reads. The parameters were set as default, and I kept only the sites more than 80 reads.)

Q2: I also utilized DENA, another m6A prediction software, and found that while both tools could identify the defective phenotypes in the 3'UTR between the mutant line and the wild type, the predicted ratios were significantly different. Could you provide some insight into why this might be?

Take WT for example:

Q3: Based on your paper, it's evident that the more reads that pass the read-level threshold, the higher the probability that the site-level could be. However, I'm struggling to understand why the read-level probability threshold is so minimal. For instance, the suggested value is only 0.033379376 for the Human model and 0.0032978046219796 for the Arabidopsis model. If a read only has a 0.01 probability of being considered as modified, does that imply that the read has a 0.99 probability of being considered unmodified?

Here is the probability distribution of all my reads in one treatment. The red line is the suggested probability threshold, 0.0032978046219796, for Arabidopsis

Q4: Both m6anet and DENA use an older version of Guppy for base calling, but improvements have been made to the calling models for the R9.4.1 flowcell from Guppy version 3.3 onwards. Could these updated models affect the accuracy of m6A prediction in m6anet?

Welcome any corrections, and I could provide more information if needed.

Thank you for your time and support.

Best regards,

YCCHEN