Inconsistent input data in scPoli tutorial: normalized values instead of raw counts

[zerostwo]

Hi scArches team,

Thank you for the excellent work on scPoli and the supporting tutorials!

I noticed a potential issue in the scPoli surgery tutorial. Specifically, the tutorial downloads pancreas.h5ad from Figshare (link) as input data. However, upon inspection, the .X matrix in this AnnData object contains normalized or log-transformed values, not raw UMI counts.

This conflicts with the expectation of scPoli, which uses recon_loss='nb' by default and requires raw count data for accurate modeling of gene expression under a negative binomial distribution. Using normalized data here could compromise both the training and downstream integration quality.

Suggestions:

• Could you clarify whether the model is still expected to work with normalized data in this tutorial case?
• If not, please consider either:
• Updating the dataset to include raw counts, or
• Including a note in the documentation that explicitly warns users that the data is normalized and should be replaced with raw counts before model training.

Thanks again for providing such powerful tools to the community!

[chen-soph]

Following up with a similar question! The model and resources have been extremely helpful and useful so far, but I wanted to also ask about using raw versus normalized counts for running and training an scPoli model.

• Adding a comment that it wasn't entirely clear whether the tutorial used normalized or raw counts. My own dataset seemed to have normalized values in the .X matrix, which is (to my current understanding) what the model draws from by default. When I ran the model with both the normalized matrix and the raw counts data, I got quite different results.
• Is the best way to change the matrix used just to reassign the .X matrix? Adding a data_layer or layer argument to the model could be helpful for future use down the line.

Thank you for creating and maintaining scPoli and tools like these!

[halterc]

I was in direct contact with the author by email and he confirmed that "raw counts [should be used] as input to the model"