Question about perturbation direction encoding and interpretation of infer outputs

[rogercasalsfr]

Hi, thanks for releasing STATE.

I am using preprocess_train, train, and then infer, and I have two brief questions.

First, how should perturbation direction be encoded? For example, if my dataset contains activation and knowckdown, does STATE distinguish them only through the labels in pert_col? Should direction be explicitly included in the perturbation name?
And if it's just one direction, it just assumes all data goes in the same perturbation direction?

Second, how should I interpret the outputs of infer?

Wrote predictions to adata.obsm['X_hvg']
Saved count predictions to adata.X

Are these predicted counts, normalized values, log-transformed values, or model-space outputs?

For example, adata.X contains values like:

 [0.         0.         ... 0.         0.        ]
 [0.13628182 0.         ... 0.11327718 0.        ]]```

But these are not counts. What should I do? Is there anything that I am missing?