Converting DeepHiC output to pFitHiC input

[smurthy55]

Hi! Hope you are doing well. I have npz file outputs from the DeepHiC pipeline using my own data. I wanted to ask more about how you converted this output to the input needed for the pFitHiC method. It seems that based on the FitHiC inputs, there needs to be a fragments file and interactions file. Could you explain specifically how you generated this interactions file? Did you use the hicpro2fithic.py script and create a bed file based on the binning of the npz file? It seems that the matrix file is organized differently from the npz files, so I wanted to know if there is perhaps another method which you used instead.

I also wanted to ask from the DeepHiC paper if there was a reason that 255 was used as the factor to convert the DeepHiC outputs to FitHiC inputs?

Thanks.

[omegahh]

Hello, smurthy55, sorry for such a long-time-delayed replying. We didn't use the hicpro2fithic.py script but using a piece of in-house script. It is very simple and only contains one function, as attached in the following, which could converting numpy 2D-array into FitHiC input.

def locus_cvter(i, resolution):
    return i * resolution + resolution // 2

def fithic_input(countsfile, fragsfile, mat, chrn, resolution, compact, bounds):
    locus = partial(locus_cvter, resolution=resolution)
    nrows, ncols = mat.shape
    lb, ub = bounds

    counts_f = gzip.open(countsfile, 'wt')
    frags_f  = gzip.open(fragsfile, 'wt')
    print(f'Writing to: {countsfile} and {os.path.basename(fragsfile)} with bounds ({lb}, {ub}]')

    maps = np.zeros(nrows, dtype=np.int)
    maps[compact] = 1
    frag_lines = '\n'.join([f'{chrn}\t0\t{locus(i)}\t{maps[i]}\t0' for i in range(nrows)])
    frags_f.write(frag_lines)
    frags_f.close()

    lower = lambda x: min(ncols, x+lb+1)
    upper = lambda x: min(ncols, x+ub+1)
    contact_lines = '\n'.join([f'{chrn}\t{locus(i)}\t{chrn}\t{locus(j)}\t{mat[i,j]}' for i in range(nrows) for j in range(lower(i), upper(i))])
    counts_f.write(contact_lines)
    counts_f.close()

For question two, we keep the DeepHiC output ranging from 0-1 unless it is required to be converted to integer. For example, like Fit-Hi-C, integral counts are required for input. We can convert them to 0-255 by multiplying 255. It makes sense that DeepHiC is trained to fit the contact map ranges from 0-255 before min-max normalization.