How to decide the two sample are the same in trancription level?

[zinky]

Hi, guys.
Is there any criterion to help us to draw a conclusion that two samples are the same or unchanged in trancription level based on RNA-seq data? Correlation test ? or No differential gene indentified ?

[fanli]

Do you mean absolute or relative transcription levels? If absolute, then you'd need an external spike-in like the ERCCs. Relative, I'd say the lack of differentially expressed genes as even in a normal two-condition comparison the correlation values tend to be very high (>0.95).

[zinky]

Do you mean absolute or relative transcription levels? If absolute, then you'd need an external spike-in like the ERCCs. Relative, I'd say the lack of differentially expressed genes as even in a normal two-condition comparison the correlation values tend to be very high (>0.95).

thanks for your suggestion. yes i mean relative transcripts level, because the library size may result in differ counts number( PS: there are alseo some other variables). After normalization, my data show correlation test << 0.05, and the correlation values is 0.97, but there are still some(300-500) DE genes based on Foldchange (abs(threshold) >2). How to explan this ?
thanks!

[amitm]

hi zinky,
You say based on fold change. Did you have replicates? If yes then maybe you can apply a standard statistical testing like a t-test.
If you are not expecting DE and seeing some (by fold change) then those might be due to sample to sample variation.
A t-test would consider standard dev. and then give p-values

[zinky]

hi amitm, i have no replicates but only two samples here. I know the way to detect DE features while my purpose is to demonstrate that there is no difference between them. You advice is useful for understanding but may not work for making a conclusion.
thanks

[fanli]

Are you getting 300-500 DE genes from a pooled variance model (e.g. DESeq or edgeR) with no replicates? Or is that based solely on the fold change?

I'd be really surprised if you managed to get 300 hits with no replicates even between two very different cell types.

[SNPsaurus]

hi zinky,
You say based on fold change. Did you have replicates? If yes then maybe you can apply a standard statistical testing like a t-test.
If you are not expecting DE and seeing some (by fold change) then those might be due to sample to sample variation.
A t-test would consider standard dev. and then give p-values

Care should be taken not to interpret a t-test p-value that is above some threshold like 0.05 to mean that the samples are the same. The samples could, for instance, be very different but the experiment lacks the statistical power to detect the differences, or the variance is too high (as you say). But the opposite of differently expressed with statistical significance is not similarly expressed with statistical significance, it is "not differently expressed with statistical significance". There are tests that try to measure statistically if two distributions are the same, and this was the basis of a flurry of reports on problems with how statistics are used in many clinical studies, but the exact tests escape me!

[zinky]

Are you getting 300-500 DE genes from a pooled variance model (e.g. DESeq or edgeR) with no replicates? Or is that based solely on the fold change?

I'd be really surprised if you managed to get 300 hits with no replicates even between two very different cell types.

yes, FoldChange is taken in consideration mainly. I also performed a test which was proposed by Audic S et al in 1997(ref:Audic S, Claverie JM. The significance of digital gene expression profiles. Genome Res. 1997;7:986–995. ). This statistic
model combine the bayes and poisson distribution, it finally gives a p value to help screening DE features(fdr was also calculated). So based on this method, i got 300-500 DE genes while the correlation of two samples transcript profile is bigger than 0.97.