mean vs. geomean

[grst]

Hi @ebecht and @FPetitprez,

I've noticed some inconsistency between the code and the paper and I'm wondering what was the original intention.

The paper states:

Given a set of transcriptomic markers of a given category, we computed a corresponding per-sample score, called hereafter a MCP-counter score, using the log2 geometric mean of this set of markers.

While the implementation just calculates an arithmetic mean:

MCPcounter/Source/R/MCPcounter.R

Line 14 in b6eac73

apply(xp[intersect(row.names(xp),x),,drop=F],2,mean,na.rm=T)

Now if the input data were log2-transformed it would be somewhat like the geometric mean, but also not precisely, because the geometric mean would require an exp() after the arithmetic means of the logarithm.

geomean = exp(mean(log(X)))

---

I'm mostly asking because in immunedeconv we recommend the users to specify raw TPM and forward them to MCPcounter unchanged, because I was assuming that it calculates a geometric mean internally. However, given the actual implementation I think it would be more appropriate to log1p transform TPM values first to not give disproportional weight to more highly expressed genes.

[ebecht]

Hi @grst

Thank you for your message and all the work on immunedeconv, I'm often referring people who wants to use deconvolution methods to it.

You're right the paper is unclear on that part, what I meant is that we compute the log(geometric mean). We indeed expect inputs to MCPcounter to be log-transformed.

As you said the goal was to equalize weights across genes by stabilizing variance for highly expressed genes. This was developed on microarray data when log-transformed intensities were the norm.

I agree with you that it'd be better to log1p transform the TPMs first !

I hope it clarifies things.
All the best
Etienne