The scrna_raw_counts_properties method has an error in the calculation method of the raw_normalized_mean2 field, causing the raw_normalized_mean2 indicator to be invalid. about scvi-tools HOT 4 CLOSED

Hi, thank you for reporting it. First question: I assume correctly that you don’t get incorrect values but you get a ValueError as you have a shape mismatch. This is happening during construction of the pandas.DataFrame. Can you copy the exact error message? It should happen for raw_normalized_mean but also the two other means in the function (or are your size factors zero?)

from scvi-tools.

RuntimeWarning: divide by zero encountered in divide
scaling_factor = 1 / np.asarray(data.sum(axis=1)).ravel().reshape(-1, 1)

I think np.divide(np.asarray(data.sum(axis=1)).ravel().reshape(-1, 1), 1) should be used to replace 1 / np.asarray(data.sum(axis=1 )).ravel().reshape(-1, 1) to avoid this error

from scvi-tools.

This means that your data contains cells without any count. Is it meaningful for your model to input cells without counts (I assume it’s due to your highly variable gene selection)? I would highly recommend to filter to min_counts before training the model as otherwise training might be unstable.
@martinkim0 We could add a check and print a warning when setting up a model if cells contain less than let’s say 30 counts.

from scvi-tools.

Thanks

from scvi-tools.