Modes for transition matrix calculation.

[EmilyMiddendorp]

Is there documentation on the modes available to calculate the transitions matrix?
I am trying to decide which mode to use to analyse a dataset where we expect to find plastic event (dedifferentiation). So at the end I am interested on the absorption probabilities between the (biologically known) terminal states towards the initial state.

[michalk8]

I am trying to decide which mode to use to analyse a dataset where we expect to find plastic event (dedifferentiation).

I am assuming you're referring to the VelocityKernel. The modes are explained in https://cellrank.readthedocs.io/en/stable/api/cellrank.tl.kernels.VelocityKernel.compute_transition_matrix.html#cellrank.tl.kernels.VelocityKernel.compute_transition_matrix
Briefly:

• deterministic - similarity between RNA velocity and connectivities (based on scheme, default is correlation, can be cosine, ...)
• sampling - sample 1 vector from normal distribution with velocity mean/var
• monte_carlo - same as above, but with many samples that are averaged
• stochastic - 2nd order approximation, see eq. 30 in our preprint

For the backward velocities (relevant to your use case), only deterministic mode is usable and the options are:

• negate - instead of calculating the similarity of velocity vector v with connectivities c, calculates sim(-v, c)
• transpose - instead of calculating transition from a cell to its neighbors, do it vice-versa

There are also other approaches that don't require RNA velocity:

• if you have meaningful pseudotime values, you can also try inverting PseudotimeKernel/CytoTRACEKernel as:

pk = ~PseudotimeKernel(adata, time_key=...)
pk.compute_transition_matrix(...)

• or if you have time-series info (e.g. days), I'd recommend taking look at WOTKernel, here's the tutorial

Lastly, if you're using the GPCCA estimator in the downstream analysis, sometimes, the macrostates can also correspond to initial states (in the forward case; this can bee seen in our CytoTRACE tutorial, cell 17: https://cellrank.readthedocs.io/en/stable/beyond_rna_velocity.html).
The GPCCA estimator has 2 private methods that you can use to export these to the AnnData object, if needed:

GPCCA._compute_initial_states
GPCCA._set_initial_states_from_macrostates