Dear list,

I have a microarray experiment with several samples grouped as: control, drug_conc_a, drug_conc_b, drug_conc_c. That is, a control group and then other groups with different concentrations of a drug (same drug): a, b and c

Besides the comparisons of the control group with respect to each of the concentrations, I'd like to study or compare the three concentrations to highlight potential trends. What is the best way (i.e. most informative and appropriate) to do it? Possibilities for limma:

drug_conc_a - drug_conc_b

drug_conc_a - drug_conc_c

drug_conc_b - drug_conc_c

OR

drug_conc_a - (drug_conc_b - drug_conc_c)

OR

(drug_conc_a - drug_conc_b) - drug_conc_c

OR

(drug_conc_a - drug_conc_b) - (drug_conc_b - drug_conc_c)

Not sure which comparison is the proper one.

Thanks,

Javier

Well, it depends on what kind of trends you're interested in. The simplest approach is to look for genes that go up or down with increasing concentration, in which case I would do something like:

con <- makeContrasts(concLow - control, concMed - concLow, 
    concHigh - concMed, levels=design)
fit2 <- contrasts.fit(fit, con)
fit2 <- eBayes(fit2)
res <- topTable(fit2)

This will do an ANOVA-like analysis looking for any effects of drug treatment. I would then restrict myself to the significant genes with the same sign for all three log-fold changes, to identify genes that responded in a consistent direction to increasing concentration. If you want to be more rigorous, then you can do each of these pairwise comparisons separately (see coef in topTable) and then intersect the sets of significantly up- or down-regulated genes across the three comparisons.

If you're looking for more specific trends, then you'll have to use a bit more finesse. For example, to find genes where the effect of the drug plateaus at a certain concentration, you'd look for large log-fold changes/strong DE up for all comparisons up to the target concentration, and then look for the absence of changes for all comparisons involving higher concentrations.

P.S. Three concentrations are probably not enough to fit a meaningful covariate trend, so I would stick with a one-way layout.