Is it possible to transfer kernels and estimators across datasets?

[dawe]

Hi all, I am analyzing a scRNA-seq dataset with cellrank. Results are good or, at least, interesting.
Now, for the same cells I could generate a count matrix on totally different features and I would like to, at least, retrieve the lineage drivers from this new features using the estimates made on RNA. I can transfer from one AnnData to the other all the informations stored in .obs, .obsm and possibly .uns, but it seems that I have to recompute a new kernel and a new estimator.
I have tried to substitute the adata reference in the estimator (g_fwd.adata = new_adata) but compute_lineage_drivers() complains about different sizes in features (which is kind of expected...).
I am now looking at the source code, I may try to call cellrank functions directly, but any suggestion is very welcome.
Anyhow, what would be the best way to transfer kernels and estimators on different views of the same data? This may be relevant when multimodal data are available (not my case, though).

[dawe]

I've found that calling cr.tl._utils._correlation_test() directly works for me, given that data is the new AnnData and lin_probs is the estimator .absorption_probabilities.

[Marius1311]

Nice, does that solve your issue?

[dawe]

It does. Now I only have to figure out if results make any sense, but that is a completely different story 😄