moscot.problems.generic.SinkhornProblem.prepare

SinkhornProblem.prepare(key, joint_attr=None, policy='sequential', cost='sq_euclidean', cost_kwargs=mappingproxy({}), a=None, b=None, xy_callback=None, xy_callback_kwargs=mappingproxy({}), subset=None, reference=None)

Prepare the individual linear subproblems.

See also

• See Custom cost matrices on how to pass custom cost matrices.

• TODO(michalk8): add an example that shows how to pass different costs (with kwargs).

Parameters:

• key (str) – Key in obs for the SubsetPolicy.

• joint_attr (Union[str, Mapping[str, Any], None]) –

  How to get the data for the linear term:

  • None - PCA on X is computed.

  • str - key in obsm where the data is stored.

  • dict - it should contain 'attr' and 'key', the attribute and key in AnnData, and optionally 'tag' from the tags.

  By default, tag = 'point_cloud' is used.

• policy (Literal['sequential', 'explicit', 'star']) –

  Rule which defines how to construct the subproblems using obs['{key}']. Valid options are:

  • 'sequential' - align subsequent categories.

  • 'explicit' - explicit sequence of subsets passed via subset = [(b3, b0), ...].

• cost (Literal['euclidean', 'sq_euclidean', 'cosine', 'pnorm_p', 'sq_pnorm', 'geodesic']) –

  Cost function to use. Valid options are:

  • str - name of the cost function, see get_available_costs().

  • dict - a dictionary with the following keys and values:

    • 'xy' - cost function for the linear term, same as above.

• cost_kwargs (Union[Mapping[str, Any], Mapping[Literal['x', 'y', 'xy'], Mapping[str, Any]]]) – Keyword arguments for the BaseCost or any backend-specific cost.

• a (Union[bool, str, None]) –

  Source marginals. Valid options are:

  • str - key in obs where the source marginals are stored.

  • bool - if True, estimate the marginals, otherwise use uniform marginals.

  • None - uniform marginals.

• b (Union[bool, str, None]) –

  Target marginals. Valid options are:

  • str - key in obs where the target marginals are stored.

  • bool - if True, estimate the marginals, otherwise use uniform marginals.

  • None - uniform marginals.

• xy_callback (Literal['local-pca'] | ~typing.Callable[[~typing.Literal['xy', 'x', 'y'], ~anndata._core.anndata.AnnData, ~anndata._core.anndata.AnnData | None], ~moscot.utils.tagged_array.TaggedArray | None] | None)

• xy_callback_kwargs (Mapping[str, Any])

• subset (Sequence[Tuple[K, K]] | None)

• reference (Any | None)

Return type:

SinkhornProblem[TypeVar(K, bound= Hashable), TypeVar(B, bound= OTProblem)]

Returns:

: Returns self and updates the following fields:

• problems - the prepared subproblems.

• solutions - set to an empty dict.

• stage - set to 'prepared'.

• problem_kind - set to 'linear'.