Hi, I am using GFOLD for my RNASeq data analysis as we don't have any biological replicates. GFOLD gives the e-FDR value equal to 1 for all genes and doesn't use p-value. What is the criteria then to select significant DEGs? Does the selection only on the basis of fold change with fdr=1 for all genes justifiable?
As he do not have any replicates p-values don't make any sense. Use the foldchange. But I would not be confident at all in the results..
I would echo with @NicoBxl that since you do not have any replicates the p-value will not really make much of a sense here. However I doubt even DESeq2 or edgeR will work since as far as my reading goes you need atleast 3 replicates minimum in each condition to make your DESeq2 and edgeR run effectively. So in any case you will not be able to use them. GFOLD is designed specifically for the purpose of these short comings and you should in this case rely only on FC values for filtering.From its website it is clearly written the e-FDR is always 1 without replicates so there is no point to take it into consideration.
log2fdc: log2 fold change. If no replicate is available, and -acc is T, log2 fold change is based on read counts and normalization constants.
So use a log2fdc for filtering
Those tools mainly rely on calculating a dispersion for each gene based on the biological replicates. Then these dispersions are statistically compared between each group of samples. If you don't have biological replicates, the estimation of the p-value will be skewed because you can't calculate the dispersion. An easy example of why it could be an issue is for one reason or another your unique sample is an outlier, you won't be able to see it without replicates and the final p-value won't be the reflect of the reality. That's why in case of no replicates experiments you should look only (and with caution) to the foldchange since the pvalue is not relevant. Regarding edgeR/DESeq2 from my memories, only DESeq2 can take a design without replicates. However the authors strongly advice to be careful while analyzing the results. A lot of posts about DESeq2 without replicates can be found on the support of bioconductor.
I am sorry I have never used GFold myself so I can't say more than the others. Those pipelines are all really similar in their statistical approaches. If no FDR is below your threshold, it is either because your conditions are not significantly different or because you don't have enough power to see those differences. You can look at the ranked fold-changes with caution but the best advice is to try to obtain more biological replicates.
Log in to answer this question.
