search

bowang-lab / scGPT

https://scgpt.readthedocs.io/en/latest/
MIT License
970 stars 181 forks source link

Question about CategoryValueEncoder and ContinuousValueEncoder for value encoder #221

Open cristinaperez9 opened 1 month ago

cristinaperez9 commented 1 month ago

Hello, first thanks to the creators for the good GitHub repo!

I have one quick question regarding the value encoder that should be used for binned input data. From the paper, it seems CategoryValueEncoder is more appropriate but from the tutorials Tutorial_Integration.py and Tutorial_Annotation.py it seems that both encoders can be used. Is there any preference of one encoder over the other for binned input data?

ajv012 commented 1 month ago

+1

subercui commented 1 month ago

Hi, you can use the default settings in the tutorials. We use continuous MLP as binned value encoder. In the paper, we also stated as "We use fully connected layers, denoted as embx, for the binned expression values to enhance expressivity. This choice enables the modeling of the ordinal relation of gene expression values."