proper input for marker selection about glmgampoi HOT 2 CLOSED

Hi pchiang5,

as you have correctly described there are two approaches when doing differential gene analysis:

1. Adapt the model formula and make sure that you have only the two groups that you care about in the fit
2. Fit a model with all groups and then look at the specific contrast you care about.

For the second approach, though, you have to be careful because the coefficients can have different meanings depending on the intercept. Specifically, in your second example, contrast = 'adata$obs$leiden8' measures the difference between group 8 and whatever group was chosen to represent the intercept. This explains why you found that the results between the two approaches don't match.

Which input would be a more reasonable/proper way for the purpose?

In summary, although I usually prefer the second approach because it is more flexible, the current contrast that you are using is wrong for identifying marker genes in cluster 8.

To learn more about correctly forming a model design and choosing the appropriate contrast, I recommend to take a look at:

• CoefExplainer and ExploreModelMatrix, two R packages for learning about different model formulas
• F1000 paper by the people who developed limma called a guide to creating design matrices for gene expression experiments
• The DESeq2 vignette which contains tons of examples how to perform differential expression analysis
• Or if you want to learn all the intricacies, the book _Statistical Design and Analysis of Biological Experiments
  _ contains even more explanations around this whole question.

Best,
Constantin

from glmgampoi.

Thank you for your answer and references, Constantin.

from glmgampoi.