What is the minimum of sample to train a model on?

[NicolaasVanRenne]

Would it make sense to train a model on two samples?

Is there an absolute minimal sample number that is required?

[ludvb]

I'm sorry, I somehow missed your question. I think it depends on for what purpose you are using the model. For example, if you only want to infer super-resolved expression maps, doing that in even a single sample will in many cases give good results in my experience. On the other hand, if you want to predict expression based on H&E images for which you don't have expression data, you will want to make sure that the variation in your training set matches the variation in the test data. So it's very much context dependent, and it's difficult to give a precise number here.

Two samples should be good for super resolution but probably not for predicting expression from histology in unseen samples.