Training on LINCS data

[hraeder41]

Hello, I have a quick question on how the model is trained on the LINCS dataset. Based on what I understand, the model learns how to make a "baseline" representation of the gene expression through disentanglement, and also based on the expression of the cell line under "control" conditions (i.e. treated with DMSO). However, I saw that the LINCS training dataset only contains "control" expression data for 17 of the 82 cell lines. How was the baseline expression established for the other cell lines? I.e., how could you evaluate if the predicted disentangled baseline expression was actually similar to the true latent representation of the real baseline expression?

I am asking because I am trying to apply chemCPA to another dataset with the same cell lines as LINCS, but the basal gene expression measurements are scaled slightly differently. I am trying to adjust them to match the LINCS distribution, but I'm not sure how given that only a subset of the cell lines actually had measured baseline expression in the training dataset.

Once again, thank you for your insight here, it is greatly appreciated!