Question about dataset preprocessing

[h5li]

Hi, I am also working on predicting DNA methylations from DNA sequences.

But I have a question about your CNN and dataset to be trained on. Is dataset labels binary? ie, [0,1] or an array with two values, ie, probability of unmethylated or methylated? Like [0.2,0.8]. Also, how did you preprocess those sequences with multiple reads/experiments?

I appreciate your time to answer this question.

All the best,
Han

[haoyangz]

Hi Han, thank you for your interest in our work. We binarized the labels. The details of this and how we merged different experiments can be both found in the paper.

[h5li]

Thanks for replying!
Just one more question!

May you specify which part of this paper on how you merged different experiments? I only found some in the section of "High-throughput DNA methylation data".
You said that "We merged multiple replicates for the same experiments, and where a CpG exists in all replicates we merged the counts of methylated and unmethylated reads and re-calculated the percentage of methylation."

Does it mean that with 2 methylated reads and 8 unmethylated reads, the percentage of methylation 20%, you binarized it into 0?

All the best,
Han.

[haoyangz]

@h5li Yes in this case the methylation is 20% (not 80%) and we binarized as 0. The 50% binarization cutoff is determined based on the highly bimodal distribution of methylation percentage across the CpG sites on the genome (centered at 0 and 1).

Haoyang