Hello,  I am pretty new to edgeR but have read through the documentation and I haven't found any examples that seem to match my scenario.  I'm using edgeR to analyze genome wide genetic screen data.  I had three replicates of my library, each infected independently, however, before the beginning of the screen one replicate became contaminated.  At the start of the screen, I took one of my remaining two replicates and split it into two, maintaining each separately for the duration of the experiment.  Thus, I now have two samples that represent my start time point.  One of those (Replicate A) has a single paired end point (A.end).  In Rep B, I now have two end samples (B.end.1 and B.end.2) that each pair with B.start.  

In my experiments in which all 3 replicates make it through the screen, I have generally been using "Replicate"=c(1,2,3,1,2,3) and "Timepoint"=c("Start","Start","Start","End","End","End") to create my design matrix.  Would the correct way to amend this to reflect my current situation be: "Replicate"=c(1,2,1,2,2), "Timepoint"=c("Start","Start","End","End","End")?

Thank you!

In other words, you have two populations (A and B) for which you have 1 start sample and 1 (A) or 2 (B) end samples. I suspect that it would not be appropriate to treat the split end samples for B as biological replicates of each other. In your experiment, the true biological replication is that between populations; the two end samples for B provide a weaker level of replication because they don't capture the variance between populations - obviously, given that those two samples are derived from a single population. I would guess that this would lead to underestimation of the dispersion, as the low variability between the two end samples will pull down the estimate of the variability between populations (which is what you want to model).

The simplest way to avoid this, while still using the information in the two end samples, is to pool those two samples together into a single end sample for population B. Then use a standard paired-samples design as you've done before, but this time with only two populations. This pooling approach is most sensible when the variability between the two end samples is substantially smaller (ideally Poisson, like summing technical replicates) than the variability between the two populations; otherwise, it might be safer to just drop one of the end samples altogether.