Is lfcShrink only useful only for visualization?

Question

[Fawzi Yassine]

Hi,

Michael Love (the developer of deseq2) has said:

The shrunken fold changes are useful for ranking genes by effect size and for visualization.

Does this mean that the shrunken fold changes [produced by lfcShrink ] are [only] useful for ranking genes by effect size and for visualization.

Regards,

[Fawzi Yassine]

The authors in (Lazaropoulos et al., 2015) said: ” Patients with FRDA have frataxin levels in peripheral tissues that range from 2% to 30% of control level” (they used Western blot for measurements). So I guess researchers agree on the 0.3 FC number. Add I got this, but withot shrinking in deseq2 – which seems necessary in deseq2. Lazaropoulos, M., Dong, Y., Clark, E., Greeley, N. R., Seyer, L. A., Brigatti, K. W., … Lynch, D. R. (2015). Frataxin levels in peripheral tissue in Friedreich ataxia. Annals of Clinical and Translational Neurology, 2(8), 831–42. https://doi.org/10.1002/acn3.225

[ATpoint]

Western blot measures protein abundance, so you are comparing apples with peers.

Answer 1

[Devon Ryan]

No, which is why he didn't use the word "only". Shrunken fold-changes are generally useful since they're (A) more conservative and (B) will tend to match what you'd see with large sample sizes (if you had them).

[Fawzi Yassine]

Thank you for your reply. Actually, I am interested in (Fold Change) FC for a single gene FXN. For the biological condition I am testing, it has been shown (by microarray) that the FC for FXN is 0.3. In deseq2, the unshrunken FC is exactly 0.3 whereas the shrunken one is 0.45 That’s why I am reluctant to use the shrunken one. Regards,

[Friederike]

Microarrays and RNA-seq generate very different types of measurements, i.e., fluorescence intensities following hybridizations to pre-selected target regions vs. counting of randomly samples regions, respectively. You should absolutely not expect to end up with the same values. It's like expecting the same values for weighing a gallon of water and measuring the time it takes to evaporate a gallon of water.

[Fawzi Yassine]

The authors in (Lazaropoulos et al., 2015) said: ” Patients with FRDA have frataxin levels in peripheral tissues that range from 2% to 30% of control level” . So I guess researchers agree on the 0.3 FC number. Add I got this, but withot shrinking in deseq2 – which seems necessary in deseq2. Lazaropoulos, M., Dong, Y., Clark, E., Greeley, N. R., Seyer, L. A., Brigatti, K. W., … Lynch, D. R. (2015). Frataxin levels in peripheral tissue in Friedreich ataxia. Annals of Clinical and Translational Neurology, 2(8), 831–42. https://doi.org/10.1002/acn3.225

[Friederike]

You're right, logFC values may be a bit more stable across different platforms than the actual values used as proxies for expression. You're lucky that for that particular gene the fold changes you measured seem to match with the Western Blot measurements of proteins (!), but it certainly is no argument against the general applicability of shrunken logFC values (even if I was to take the values at face value, I would have no problem accepting that there's a 37% increase of the transcripts). RNA-seq is not a super quantitative assay (neither are Western Blots!) unless you used very carefully selected spike-in controls that actually allow you to gauge the slope of recovery for transcripts with similar properties as the one you're interested in.

I guess, the main message everyone here is trying to bring across: do not take either logFC result at face value. I'm sure I could produce a (slightly) different one by using the same data but different annotations, transcriptome references, or different tools for calculating the logFC.

[Fawzi Yassine]

Hi Fred, So which one do you command that I report? The unshrunken LFC or the shrunken one? By the way that disease I am experimenting with end which I carry is called Friedreich’s Ataxia.

[Friederike]

Hi Fa, I don't think it matters as long as you stay consistent. As Devon pointed out, the shrunken logFC is assumed to be a better approximation, but there's no guarantee, of course, because it always depends on the noise, experimental design etc. of your samples.

[Fawzi Yassine]

lfcShrink_res_deseq2 = lfcShrink(dds=dds, coef = 2, res = res_deseq2, type = "apeglm") gave 0.32 FC which is perfect!