0
$\begingroup$

I have a dataset that looks something like this

Treatment1 Treatment 2 Control
Sample1 3.23. 0.87. 1
Sample2 1.71. 1. 1
Sample3 2.88. 5.65 1
Sample4 0.77. 1.34 1

The numbers describe a fold change with respect to the control (where the control is always set to 1).

Does it make sense to compute values such as averages, sums, and differences (e.g. the average fold change for treatment1)? And yes, wary that doing things on log fold-change is not a thing, but how about just fold change?

By extension, does it also make sense to compute something like Euclidean distances (e.g. fold change vector of Sample1 vs. Sample2) for clustering fold change data? I think clustering on counts/TPM makes more sense, but wasn't sure about doing it on fold changes.

Something tells me that it's not possible because the fold changes are relative values as opposed to measured numbers. I've opted for looking at correlations (e.g. Spearman) for now though; are there alternative metrics/strategies to consider?

If there are some good papers talking about clustering and manipulating fold change data I'd be equally grateful, thanks!

$\endgroup$
2
  • $\begingroup$ Your control is 1? How did that happen? Would it be more straightforward to cluster the samples/genes directly on the expression data (often visualized using heatmaps). e.g.: ebi.ac.uk/training-beta/online/courses/… $\endgroup$
    – Kamil S Jaron
    Jan 29 at 15:37
  • $\begingroup$ Ah, it's been given to me where the control is deliberately set to 1 (as it's differential expression with respect to the control?) $\endgroup$ Jan 29 at 16:07
1
$\begingroup$

Does it make sense to compute values such as averages, sums, and differences (e.g. the average fold change for treatment1)? And yes, wary that doing things on log fold-change is not a thing, but how about just fold change?

This is problematic, if you want to compare treatments then compute the actual fold change and its significance. You can't really base anything off of the fold-change (log or otherwise) alone.

does it also make sense to compute something like Euclidean distances for clustering fold change data

It's not unheard of to do this. The problem is you lose all information about how reliable those fold-change estimates are. This is something maintained when you cluster sample-level information (e.g., TPM or robust log transformed data). Not that using variance-stabilized data for clustering will work better, since it tamps down on how much highly-expressed genes influence your clustering (assuming you don't want that to occur).

Have a look at the DESeq2 vignette for some nice examples of what good practices are.

$\endgroup$
1
  • $\begingroup$ Great, thanks Devon! Having come from a different branch of bioinformatics, it's not exactly clear what to do next once you have those DEGs. I agree that clustering on DEGs is, well, doable, but I can't tell if it's the "right" thing to do as I feel you're doing this on relative data (i.e. ratios) as opposed to an actual number. e.g. if we applied clustering by Euclidean distance, what does it mean when you have to subtract fold changes? That was my big debate, really... I'll have a look at that vignette!! $\endgroup$ Jan 29 at 14:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.