Handling zeros

[dawe]

Hello, I was applying propr-derived code on some scRNA-seq data and I got some results that prompted me to investigate the expression level of some interesting genes. Due to the fact some of the interesting genes in my settings are full of zeros, I wanted to check how zero are replaced in propr. According to the paper 0 is simply replaced with 1, this behaviour has been changed (I see there's a related issue #23 ) and 0 are replaced with the next smallest values. In the paper is

For simplicity, propr automatically replaces all zero values with 1 prior to log-ratio transformation, corresponding to the multiplicative replacement strategy

But reading the referenced paper (Martín-Fernández et al, 2003) the multiplicative replacement strategy requires the non-zero values to be replaced as well, multiplying them by the proportion of the non-zero values in the sample (eq. 6), to maintain the proportions.

[dawe]

I have performed some tests and it seems it's irrelevant whether data are renormalized or not after pseudocounts are added, the clr is basically the same (unless I made some gross mistake).
The question shifts to "how should zeros be handled in scRNA data, where >70% data are 0?"

[suzannejin]

Hello @dawe,
Just wondering how have you solved the situation so far?

As for the zero replacement, yes, the default option within propr replaces them by the minimun non-zero number in data. In the case of multiplicative replacement strategy, it also imputes the zeros by a small number.

As for the clr, it will return the same numbers if you simply renormalize by scaling the cells. This is because the clr is simply count divided by the cell geometric mean (if your count is larger, the geometric mean as well, so the centered logratio maintains the same, no matter the scale).

As for your main question, the choice of imputation methods often drives the analysis outcomes largely due to the amount of zeros in single cell datasets. Some recommendations are to use metacells. Not sure if you have explored already?