ChicagoTools fitDistCurve.R: how to fit p-value weights when comparing two treatments?

Question

[adrija.kalvisa]

I am trying to estimate the p-value weight settings by using ChicagoTools fitDistCurve.R using replicates from my own samples as an input. I want to compare two treatments (three replicates per treatment), however, I am not sure whether I should calculate a separate weight adjustment for each treatment.

For comparison, I include the p-weight values from the built-in .settings files:

                GM12878    humanMacrophage          mESC
weightAlpha   29.138483          34.115730     18.259997  
weightBeta    -2.342790          -2.586881     -1.547562 
weightGamma  -17.108579         -17.134780    -17.357088 
weightDelta   -7.688056          -7.076092     -7.216534

My own .settings files, calculated from my own replicates:

                  TreatmentA    TreatmebtB
weightAlpha       24.486049     23.696801
weightBeta        -1.978451     -1.931931
weightGamma      -17.135199    -18.066895
weightDelta       -6.737233     -7.222880

You can see that the weights of the p-values differ when generated for each treatment separately. I don t know whether the treatment can affect the "preferred" interaction distance (I would actually like to test this).

Taking all this into consideration, I have two questions:
1) Does it make sense to "pool" the samples from different treatments for p-value weight calculation using fitDistCurve.R when I dont know if treatment affects differentiation state/cell type?
2) What is the expected similarity between the p-value weights when comparing two samples of the same tissue, species or treatment?

Chicago vignette

[ATpoint]

Since Chicago is at BioC I suggest you post this over at their support page: https://support.bioconductor.org/

pcHiC is not super common and the user base here is therefore probably small.

Answer 1

[a11msp]

It's a great question, and I'm not sure there's just one "correct" answer.

Weights are not required for the actual differential analysis package - Chicdiff - which only uses expected mean estimates from Chicago, and computes its own p-value weights (because the hypothesis tested by Chicdiff is not the same as in Chicago).

What the weights affect is the set of Chicago regions to be tested by Chicdiff, which are called separately for each condition and then combined to define the "peak file" for Chicdiff.

So one way to approach this is call these regions with the most optimal settings for each condition (i.e., potentially different weights).

It must be said though that in practice we often use the default weights, because they are trained on a very high-powered dataset, and it may be that they are simply more precise than those computed on custom data with fewer replicates.