lmFit model datasets requirements

Question

I am very new to R and try to analyze a few expression array data.

For the gene expression analysis, we use linear fit and eBayes to calculate the data. But if I only have one sample for each condition(say, 1 control, 1 experiment), can I still use lmFit/eBayes function or just do an order for the MA result to find out the top genes. Is it because calculating co-efficiency needs at least two samples for each condition?

I read the limma package manual. It lists some examples. I noticed that in the time course experiments (page50), the case has two 0hr wt, two 0hr mu, one 6hr wt, one 6hr mu, one 24hr wt, and one 24hr mu. It did the lmFit/eBayes process. Is it because it is a time course case? If I have a time course data, which still contain one sample for each condition (say, 1 control and 1 experiment, for 0hr, 6hr, 12hr, and 24hr), is it reasonable to calculate the co-efficiency with lmFit/eBays?

Thank you very much!

Answer 1

Since the variance is 0 you will have problem in empirical Bayes smoothing (of the standard errors). I tried with a toy example and here is the error it gives:

> efit<-eBayes(fit)

Error in ebayes(fit = fit, proportion = proportion, stdev.coef.lim = stdev.coef.lim) : No residual degrees of freedom in linear model fits

Answer 2

Yes, the time course is a special case. Expression is assumed to change smoothly over time and a temporal trend is fitted using a regression. In all other designs, you need replication to detect differential expression.