How to predict on new cells and drugs that are not included in Prophet's pretraining?

[CCCC1800]

In your tutorials, the code uses drug SMILES and cell lines' names as input (as shown in the screenshot):

However, I'd like to use Prophet to predict on new transcriptomic data, which come from new cells that are not included in Prophet's trained cell lines' name. I also want to use Prophet to predict on new drugs, which are not included in Prophet's trained drugs. How to use Prophet to do it?

Thank you.

[nictru]

Hey, I have the same problem.
I have the impression that the published implementation only supports cell line identifiers that were used during the training process.

When adapting the source code, one could also potentially allow for new transcriptomic data.

But then the next problem arises: The input to the model is not the transcriptome itself, but a PCA based on it. Of course, one could simply perform a 300-dimensional PCA on the new data and ingest it, but I assume the performance will be abysmal because the PCAs were based on different raw data.
Therefore, it would also be necessary to make the loadings of the original PCA accessible, enabling users to generate representations that Prophet actually understands.

Given the amount of effort that went into training this model, it would probably be a relatively minor effort to make this possible. If the authors could make the PCA loadings available, I could try to open a PR with the necessary changes