Like that we can get cell-type fractions like with "classical" deconvolution algorithms. Consider normalizing the estimates per bulk sample to sum to one.