is it possible to define an end state when using temporal kernel ?

[AtlasMaroc]

...
Hello, we have a biological prior with verified marker genes in the lab, thus we kinda have an idea about the terminal state, however when using fit(), it doesn't appear. is there a way to define the terminal macrostate ?

[Marius1311]

Yes, please see our tutorial: https://cellrank.readthedocs.io/en/latest/notebooks/tutorials/estimators/600_initial_terminal.html#advanced-usage

There's a section where we use set_terminal_states

[AtlasMaroc]

@Marius1311 thank you for the response, i meant select a cluster or labaled cell from adata.obs column.

[Marius1311]

That's also possible, check the API: https://cellrank.readthedocs.io/en/latest/api/_autosummary/estimators/cellrank.estimators.GPCCA.html#cellrank.estimators.GPCCA.set_terminal_states

[AtlasMaroc]

That's also possible, check the API: https://cellrank.readthedocs.io/en/latest/api/_autosummary/estimators/cellrank.estimators.GPCCA.html#cellrank.estimators.GPCCA.set_terminal_states

Thank you, one additional question, i have been curating data across time point in embryonic development for a specific tissue, my concern is batch effects, would it skew my results and how i interpret the results, and how do i mitigate/navigate the issue? From what i am aware, integration methods just find a shared manifold for visualization, and don' regress out the data.

One other thing is, it is possible to get a tree like structure? One example i had in mind is the flow model from stezlerlab implemented in their mouse gastrulation paper.

[Marius1311]

I recommend reading up on the WOT paper and related work, e.g. moscot, and some of the other more recent OT papers in the field. You can get a batch-effect corrected latent space with Harmony/scVI/scANVI/scPoli/etc. , and then run OT in that corrected latent space. CellRank can work with the output of moscot/WOT though it's Realtime kernel, check our tutorial on that, and the API for the Realtime kernel.

We don't fit trees to the data, CellRank treats differentiation as a probabilistic, continuous process. See our tutorial for more info, and read the CellRank papers for background.

[AtlasMaroc]

You can get a batch-effect corrected latent space with Harmony/scVI/scANVI/scPoli/etc. , and then run OT in that corrected latent space

Thank you.

one other question is the interperation of the results, when i set the terminal state as a cluster defined by me, i get different result compared when i use the predicted terminal state, in terms of correlated genes? also what would be a good correlation threshold ?