RNA seq analysis with one replicate per biological sample

[anle]

Hi, I am new in the field of deep sequencing analysis and for the past two months I am struggling to understand RNA seq differential expression studies.
I have 1 control sample and 1 sample for each of the three different experimental conditions. No technical replicates. So I guess I cannot apply some GLM approach or estimate mean and variances from the replicates.
I chose to normalize with applying a simple linear correction after fiiting the log # of reads ( condition x vs control) and then to calculate the significance of the Differential expression by calculating a z score for each gene by using my control sample to estimate per gene mean and variance (which have practically the same values). The p values that I get are extremely low even after bonferoni correction (eg top 1000 pvalue still less 1e-100). This is quite counter intuitive for interpretation. Any other suggestions for better normalizing or studying DE?? thanks

[Cole Trapnell]

How many genes do you expect to be differentially expressed? If that number is a small fraction of the total (say 10% - others may have better guidance and experience in this situation), you can treat your conditions as replicates to model the biological variability in fragment counts as a function of expression level. Cuffdiff (1.0 and later) and I believe both DESeq and edgeR can do this.

When you give Cuffdiff samples from three conditions, and each has only a single replicate, Cuffdiff looks at the mean fragment count in each locus across replicates, calculates the variance at each locus, and fits a model of fragment variability across replicates. This same model of variability is used for each of the three conditions during differential expression testing. Because the the genuinely differentially expressed genes are going to amplify the variance, the model will predict more variance than really exists. However, if the number of differentially expressed genes is small, the overestimation of variability will be small. It's not an ideal situation, but its better than no empirical model of variability at all. Had you replicates in each condition, Cuffdiff would fit a model of variability for each condtion and use the appropriate models during testing.

So the short version of my suggestion is just to try Cuffdiff - it will do the above by default with your data.

[para_seq]

Hi, Cole,

Thank you for the explanation. Do I understand it correct that the variance value empirically estimated by cuffdiff, DESeq, and edgeR is gene-specific, i.e. different genes will have different variance values? If the answer is yes, how does the relative number of differentially expressed genes out of the total genes has any effect on how good (or bad) the test for DE would be? Does the variance value get somehow normalized across read count of all genes?

Thank you.


Ben