Hi all,
I've noticed a suspicious pattern in some volcano-plots I've made of proteomics data.
Specifically, I've noticed points lining up, seeming to show a direct mathematical relationship rather than the expected volcano shaped cloud.
I've normalised and quantified the data using the R package `MSstats`.
I've also also noticed some proteins showing a p-value of 0, I've filtered these out however I believe some people assign a very low number to replace this. What is the usual practice here?
Volcano plots attached.
Many thanks,
Sam
These plots seem to indicate a problem with normalization. Or is it expected and reasonable that almost all proteins are regulated in the same direction?
The few points lining up could be proteins fro which you only have a single measurement (and the standard errors are etszimated only from the pooled variance of all other proteins - which is a constant, leading to a strict functional relationship between LFC and p-value).
The multiple-testing adjustment introduces another layer or disturbance to a volcanoplot. It's better to show unadjusted p-values.
Classifying the proteins into up- and down-regulated based on p(adj) < 0.05 is not very wise. This implies that you had a dedicated hypothesis about the regulation of each of the proteins, what seems unrealistic. If you already use volcano plots, then identify proteins that are outlying, preferrably in the upper-left and upper-right arms.
The huge LFCs can indicate a detection problem. Such values occur easily when the expression (or the detectable amounts) are very small, and there is a signal very close to zero measure in the group used in the denominator. If this happens, minor differences in the protein preparation can already lead to extreme artefacts in the analysis. This might be the case in your samples/measurements. I have no idea what the detection limit (LoD) and the quantification limit (LoQ) of your method is. You should check how close the values are to these limits. If many proteins are close to these limits, it might be better using censoring in the analysis (e.g. Tobit or Survival models).
Thank you for your answer Jochen Wilhelm,
You've given me plenty to think about!
I've made some volcano plots with p-value as well and adjusted p-value, and the relationship is still present.
There's a chance that it could be due to missing data, as mass spectrometry has a fair bit of run to run variability. I tried filtering out proteins with rates of imputation, but the lines still persisted.
Sorry for not being clearer in this question, stats is not my strong suit!
Sam
Here is the same plot including only values with less than 10% imputation
Hi all,
I've narrowed down the cause of this phenomenon, (in case anyone else has this same problem), namely there's an issue with the dataset I'm using. Running the same analysis on a different dataset using the same script does not produce this pattern.
Notably, the pattern varies as a function of the normalisation method used in MSstats.
Sam
- Badges
- Science topic