Variation in RPKM values of different biological replicates produced by Cufflinks

[thanhhoang]

Hi all,
I want to rank gene expression based on the transcript abundance to see the relative expressions of different genes in my RNA-seq sample. I ran Cufflinks for 3 biological replicates for that condition to get the FPKM values. However, the FPKM values are very different within the 3 biological replicates, which I am not sure why and how to deal with this.
In Cufflinks manual, It said that Cufflinks can only use classic-fpkm normalization on one library at a time. Should I use Cuffdiff2 to normalize FPKMs and fragment counts across all libraries (This policy is identical to the one used in DESeq) and take the average of these normalized RPKMs?. Or do you think it is ok to get the FPKMs from biological replicates produced by Cufflinks and feed them into DESeq to normalize them?

Thanks for your opinions.
Thanh

[mbblack]

The FPKM values would not be expected to be the same or even necessarily very similar amongst biological replicates. That is the point of using biological replicates - large differences in relative transcript abundance are often due (at least in part) to innate biological variation in expression.

For some conditions/systems/instances the actual individual variation in expression may be the single largest source of variation in expression amongst the samples. For example, in human exposure to arsenic, gene expression estimates show the intra-individual variance is far greater than the inter-individual variance. PCA plots of gene expression estimates for individual humans will not show any overlap even out to 3 or 4 standard deviations - each individual's expression is largely unique (at least for some tissues).

That's the problem with trying to rank or compare individual expression estimates - you have single point estimates of expression per transcript, and those are whatever they are.