Defining the formula/contrasts to detect DEGs between pre vs post treatment

Question

[Dahham]

Dear all,

I would appreciate hearing your thoughts on a gene expression study we are conducting.

Background on the study: We aim to assess the effect of treatments (A, B, C, and D) on gene expression in whole blood. Samples were drawn from each patient at two timepoints: before treatment (T0) and at outcome assessment (T1). The outcome is either 1) a response or 2) no response. Within the outcome group of “response”, there are patients who just partially responded, and there are those who reached full remission.

Q1. To assess the treatment effect of drug A in those who responded, is assigning the formula as (~ 0 + timepoint), and defining the contrast “T1_T0 = timepointT1 - timepointT0” using limma/voom the best way to capture any possible therapy effect?

How should one account for the fact that these are “paired” samples in limma/voom? Should I instead switch to DREAM and include patients as a random effect to account for this paired sampling? (within each drug group, a patient has contributed to only 1 pre-treatment sample and 1 post-treatment sample)

I will of course also adjust for other important disease covariates and batches.

Q2. How should I define the formula/contrasts if I wanted to check for DEGs between T1 and T0 in those who responded compared to those who did not? Or in those who partially responded vs full remitters?

Q3. If I want to disentangle a possible “specific” effect of drug A from other “shared” effects that all other drugs may provoke (drugs B, C, and D), how should the formula be assigned in this case?

Q4. Are there bioinformatic frameworks other than limma/voom/dream/edgeR/DES/deseq2 that are better suited for my research questions?

Thanks in advance for your kind help

Dahham

Answer 1

[i.sudbery]

Q1

I would use the formula ~ patient + timepoint. This account for the pairing of the samples. You can then do the same contrast as you describe.

Q2 and Q3

You are looking at what the limma manual calls a "within and between" design. You can read about edgeR's approach to this on page 42 of the edgeR users manaul. I tend to take a more traditional GLM approach, and fit something like ~patient + timepoint + timepoint:treatment. Testing the interaction term will tell you things that vary between treatements.

Q4

Not that I'm aware of.

[BioinfGuru]

Just to add, this blog for help with model design has helped me alot.

[Dahham]

Thank you, both, so much! i.sudbery BioinfGuru