Q: How many fragments are generally sufficient for hmmratac function?

[leqi0001]

Hi,

I have a big cohort of scATACseq data, totaling 1 billion fragments from 24 pools of experiments (24 billion total). I've been trying to run hmmratac on these 24 fragment files with --cutoff-analysis-only, but it has been a week. The log says it downsampled to 800 million fragments for training, but the step to generate short, mono-, di-, and tri-nucleosomal signals has run for 3 days. Should I downsample the fragments files before hmmratac? There must be diminishing returns with more fragments, but I'm not sure how to decide the degree of downsampling.

Appreciate any suggestions!

[taoliu]

1. 800 millions reads is still too much for one single run on human genome... You can further down-sample to about 50million.
2. As for scATAC, you may want to select a subset of cells to call peaks.
3. Did you use MACS3 v3.0.2?

[leqi0001]

@taoliu
Thanks for your reply! I'm trying downsampling and see what happens. Is there any downside with using too many reads other than memory/time consumption, such as increased noise?

3. Yes I used v3.0.2.

[taoliu]

@leqi0001 First, it will be a waste of $ if you sample is already saturated. Secondly, it will take more time and memory to process -- especially for hmmratac. Lastly, for method calling peaks based on p-value cutoff, such as callpeak, p-value will be overly optimistic when the sample size is large. In this case, effective size (using foldchange) should be considered together with p-value cutoff.