[ ] Lincs vs Trapnell have different number of genes: ~970 genes in LINCS, 5000 genes in Trapnell. How do we deal with this? We could add another layer to the encoder & decoder for finetuning. We could also also have an output dimension of 5000 already on LINCS and just set all non-occurring genes to 0.
[ ] Dealing with the Adversarial classifier: If we decide to classify clusters while finetuning on LINCS (#22), then it might make sense to train a new adversarial classifier on Trapnell which uses the vanilla CPA BCE loss over all 188 drugs.