Conceptual question regarding norm vector rescaling

[dmalzl]

Hi,

First of all thank you for this amazing implementation.
I have a rather conceptual question regarding the code. From using this tool and from what I found how it is used in the HiCExplorer hicCorrectMatrix I found that there are two main results that can be returned:

1. a norm vector and a rescaled norm vector with kr.get_normalisation_vector(True/False)
2. a normalized matrix with rescaled and nonrescaled norm vector with kr.get_normalised_matrix(True/False)

My question now is: Why the rescaling?

I get that the nonrescaled results balances the matrix to rowsum/colsum of 1, but is it better to use the rescaled result instead of the unrescaled?

Also a line in the hicCorrectMatrix script is a little bit misleading in this sense:

732            # set it to False since the vector is already normalised
733            # with the previous True
734            # correction_factors = np.true_divide(1, kr.get_normalisation_vector(False).todense())
735            correction_factors = kr.get_normalisation_vector(False).todense()

However, there is no previous True. I mean for the h5 format it does not matter since you anyway store the normalised rescaled matrix but if you use it in cooler this will get you the nonrescaled vector or am I wrong here?

Thank you for the answer in advance,
Best regards,
Daniel

[dmalzl]

Just another question:

Could you also elaborate on the way you compute the rescaling. Why do you use the square root of matrix sum ratio. Is there a deeper reason for that?

[LeilyR]

Hi Daniel,

Thanks for your interest in this tool. In theory using the re-scaled one or the original matrix generated directly by kr algorithm (with row and column close to 1) should not change your analysis, however our motivation was scaling up the values to avoid the complication caused very small values in the downstream analysis when using hicexplorer. I hope it helps, let me know if you any further questions.

[dmalzl]

I see. I also thought this shouldn't matter. Anyway, thanks for the clarification :)