In a dataset I am analysing, there is a single replicate of RNA seq (or microarray) data each for tumour tissue and adjacent non-tumour tissue. Such data is available for say 10 patients in the dataset. How do we identify DEGs in this dataset? Should I treat all the non-tumour samples as control and all the tumour samples as test considering the patients are different in terms of age, sex disease state etc? What kind of statistical test is most acceptable?
Any information would be appreciated
Thanks
Thanks for getting back to me Muhammad.
I understand it would have been better to have multiple replicates, but unfortunately the dataset I am using provides only one replicate for each tumor and tumor adjacent sample. I think the reason for this, at least in the case of liver tissues is that biopsies can be quite difficult and taking multiple samples increases the risk of compromising the health of the patient.
Hi Shivaksh,
I think you're making a confusion between technical (many measurements of the same sample) and biological (many samples of the same biological condition) replicates. If I understand correctly, you have 10 patients samples with tumour and adjacent non-tumour tissue, right? In that case, you can consider (even if you'll have an intrinsic variability due to patients heterogeneity) that you have 10 replicates of each condition (tumour vs non-tumour) and apply either the Geo2R or any other R package dedicated to microarrays analysis (like limma) or DESeq2 if it's RNA-seq. Since your samples are paired (tumour and non-tumour from the same patients), it have to be taken account for, and will increase the statistical power of your tests, so be sure to include this parameter when testing. Good luck!
You have e.g. 10 replicates of within-patient comparison of diseased-vs-healthy tissue, hence you should use a paired t-test or, if you prefer non-parametric methods), a Wilcoxon test. In a software like limma, edgeR or DESeq2, the corresponding model should look like:
~ patient + disease
The important thing is NOT to pool all controls etc but to compare within patient to adjust for patient differences.
Hope that helps.
Thanks for your reply Alex.
I understand your suggestion to get a p-value for each. Additionally, I also want to understand how to get the LogFC. Keeping your suggestion in mind, should I generate a LogFC for each gene for each patient sample pair (tumour vs non tumour) and then get a mean for all the Fold changes? or is some other method more suitable. Any information would be appreciated.
Thanks
Sorry Shivaksh, must be a misunderstanding. I'm saying the same as Fabrice. I havent't seen his reply for some reason or I wouldn't have replied at all. That is:
You have 10 replicates of within-patient comparison of diseased and non-diseased tissue. Any appropriate software can compute ONE p-value and logFC for you, not 10. (Taking the mean or some other summary of within-patient logFCs is what happens internally but you won't have to care.)
Hi Shivaksh :) in a simple words what I'd do is:
1.) calculate logFC of genes for each patient (log10 of the expression of given gene in tummor tissue / expression of this gene in non-tumor tissue)
2.) For each gene apply statistical test which will tell you if the change is significant (your null hypothesis would be that in reality the FC equals 0) - you don't have to calculate mean yourself, software should do it for you .
You can also perform the test specyfying groups (case group and control group) and test on the basis of case vs. control but this is also something that software will calculate for you, you just need to specify sample status.
As mentioned above you need to keep in mind that samples are coming from the same patient so they're paired and you need to choose the appropriate version of test
- Badges
- Science topic