More than 10 replicates

[LinearParadox]

Hi all,

I had a question regarding what is best if you have more than the max number of replicates. For example, one of the conditions in my experiment has 14 biological replicates. My initial thought was the split it into batches of 5 or so, and run these as separate runs of the pipeline. Would this work? Or is there a better way to go about this?

[akundaje]

The ENCODE pipeline was not built for this kind of design with large numbers of replicates. It is primarily for < 5 replicates. A simple hack would be to pool multiple replicates into 2 or 3 meta-replicates such that they are reasonably balanced in terms of read depth and then run them through the pipeline as if they were the replicates. But instead I would suggest using an approach that is better designed for this problem e.g. https://www.sciencedirect.com/science/article/pii/S0888754321001531

…

-Anshul.

On Fri, Aug 11, 2023 at 9:12 PM Alexei Martsinkovskiy < ***@***.***> wrote: Hi all, I had a question regarding what is best if you have more than the max number of replicates. For example, one of the conditions in my experiment has 14 biological replicates. My initial thought was the split it into batches of 5 or so, and run these as separate runs of the pipeline. Would this work? Or is there a better way to go about this? — Reply to this email directly, view it on GitHub <#422>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AABDWEJ63R6AIPMHGZVJGZ3XU37CPANCNFSM6AAAAAA3NVRO6I> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[LinearParadox]

Thank you, this was tremendously helpful. My question is if there's any analytical methods you know of that work well with both small and large numbers of replicates? A lot of what I've found seems to be optimized for one or the other. For example, one of my conditions has 2 samples, and the other has 14, and id preferably want to treat them with the same tool to make direct comparisons in peak differences.

[akundaje]

The paper I linked to claims to work for small and large numbers of replicates. Let me know if you try and if it works well.