I am performing differential expression analysis using the limma package on DSP GeoMx data, and I have a question regarding the handling of multiple ROIs per core. Below is a snippet of my sample information:

ROIs    Class   Group
A1  pos Pre
A1  neg Pre
A1  neg Pre
B1  pos Post
B1  pos Post
C1  neg Post
C1  pos Post
D1  neg Pre
D1  neg Pre
E1  neg Post
E1  pos Post
F1  pos Pre

The unique patient IDs are A1, B1, C1, D1, E1, and F1.

I have the following questions:

1. How does limma account for the fact that multiple ROIs have been selected from the same core? These ROIs are not biological replicates, and I want to ensure that this assumption is properly handled in the analysis.
2. How can I adjust for the subsampling per tissue, considering that the multiple regions of interest placed per tissue section are not independent observations?
3. How do I create a design formula to incorporate both the Class and Group variables for the limma analysis?

Thank you in advance for your assistance. Gordon Smyth

Mike Love is the author of DESeq2, not limma (which was authored by Gordon Smyth), so you should not expect Mike to opine.

1. You can use duplicateCorrelation to estimate the within-subject correlation and then fit a GLS model 'lmFit` will do this automatically if you supply 'block' and 'correlation' arguments.
2. Is this different from your first question? Do you have multiple slices per tissue and multiple ROIs per slice? The limma package only fits a GLS, and only with a global estimate of the within-block correlation (where by 'block' I mean 'samples that are assumed to be correlated'). You could assume that the ROIs are more highly correlated than the individual tissue slices, in which case you might need something more sophisticated than limma, for example nlme or lme4. But that's off-topic for this site unless you are planning to use the variancePartition package.
3. Depends on your model, and how you want to partition the variance. If you use limma and a GLS model, you could just make a combination of the two and fit a cell means model. You would then construct the appropriate contrasts to test the hypothesis of interest.