Hej,
We have performed a TMT-10plex MS experiment (10 samples divided over 2 groups) to identify differently expressed proteins. We are discussing very much how to analyse the data.
What we do now (and what I believe is commonly done) is:
1. Relative quantification based on the label-reporter ions (sample 1 is set as 100% for scaling of each scan)
2. Mascot search for ID of proteins (using Proteome Discoverer)
3. QC: score >30, unique peptides >2, scan count>3
4. Median value of all unique peptides from a protein is taken as a "protein value"). All values of sample 1 are "1", as a consequence of the scaling.
5. Normalization to the median of all proteins in 1 sample
6. Statistics (T-test or Mann-Whitney U (due to low sample number, one cannot be sure about the normal distribution)).
According to a statistician, we are violating the "independent samples" assumption to perform a T-test. This is supposed to happen in step 1, where all absolute values are divided by sample 1 to get the absolute values of different scans within the same scale.
I cannot seem to find this in (proteome) literature. Could someone help us out? Are we violating the statistical assumptions or are we allowed to continue?
Thank you,
Dear Veerle,
T-test with capital letter refers to Wilcoxon's sign test and t-test to student t-test in Scandinavia. What is your sample size? In case it is greater than (> 30) you can "violate" the assumptions of normal distribution. However, I am sure some guys are soon going to elaborate your question in details.
I don't know anything about TMT-10plex experiments, but just about the statistics, if you compute the 100% scaling value with the sample 1 of group 1 and again for sample 1 of group 2, you are indeed creating dependence of your two groups, as they will be artificially reduced to similar values (around 100%).
But if you are computing the 100% scaling value just with sample 1 of group 1 and that you use it to scale the values of all samples of the TWO groups, then I don't see why you would be violating the independent samples assumption, as you are just reducing your data but not centering them, so working on proportions instead of raw values should not change anything.
But as I don't know about your topic, I can't confirm !
Hej,
Thank you for your answer. We have 5 mice in both groups (5 wildtype vs 5 knockout). I know from literature that most people apply Student t-test for this kind of data, but our statistician as very convinced it is not right.
Indeed, a t-test on 2 x 5 data point seems not right as you may not have truly normally distributed data in both groups. But you can apply a Mann-Whitney U test if, as I said above, you are scaling the 10 individuals with a single on, but not if it's with one of each group.
Why not apply Wilcoxon's sign test (T) then? Here you can do it online:
http://vassarstats.net/wilcoxon.html
The Wilcoxon's sign test is for paired samples only, while it's not the case there if I'm correct. Comparing paired samples would be, for example, comparing sample 1 of group 1 and sample 1 of group 2 over time, while here there it's several different samples in each group.
Depending on the program you use, but running it without the "paired" parameter is actually doing a Wilcoxon rank-sum test, which is the same as Mann-Whitney U test. Yes, it's a bit confusing !
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics