"use_observed_batches" in "model.differential_expression depends on batch_correction

[MaximilianNuber]

Dear all, dear Dr. Boyeau,

Thank you for the great toolkit, it is exactly what I need to analyse my data.
I have been interested in the lvm-de, as it not only uses the scvi models I already trained, but believe that the bayesian approach with optional permutation is perfect for the heterogeneity of my data.

Specifically, I am looking to compare the expression of two groups within a cell type, a scenario covered in the respective publication.
As I have several patients per group, I typically use the batch_correction option.

However, I tried to use use_observed = True for conditioning the expression per cell on the respective batch/sample, so I would get a sample-based comparison of expression values.
This gave me the error:
TypeError: scvi.model.base._differential.DifferentialComputation.get_bayes_factors() got multiple values for keyword argument 'use_observed_batches'

Digging a bit into the source code, I found that in scvi-tools / scvi / model / base / _utils.py
the _de_core function calls DifferentialComputation.get_bayes_factors (line 250).
In this function call, the option is set as use_observed_batches=not batch_correction .

So to my understanding, if I understand correctl, whenever batch_correction is True, I cannot condition on batches, and th other way around.

This does seem deliberate, so my question is how batch_correction and conditioning on batches relate to each other?
Why does one exclude the other, or is even the opposite of the other?
Is there a way to not condition on batch, and not batch_correct at the same time?

I am trying to get the background, as I am relatively new to the field and will publish the results, for which I like to have at least a most basic understanding.

Thank you for any help.
Best,
Max