Synthetic data

Data drawn on a known tree, so a reconstruction can be scored against the answer with robinson_foulds.

datasets

Synthetic data on a known tree, for testing and for the docs.

simulate

simulate(
    n_leaves: int = 64,
    n_features: int = 200,
    *,
    kind: Literal[
        "binary", "random_branches", "unbalanced"
    ] = "binary",
    branch_length: float = 1.0,
    noise_sd: float = 0.1,
    noise_spread: float = 2.0,
    seed: int = 0,
) -> SimulatedData

Brownian motion on a known tree, with per-cell per-feature error bars.

Parameters:

Name	Type	Description	Default
n_leaves	int	Cells. A power of two for "binary" and "random_branches".	64
n_features	int	Features.	200
kind	Literal['binary', 'random_branches', 'unbalanced']	"binary" (balanced, equal branches), "random_branches" (balanced, log-uniform branches on [0.5, 2]) or "unbalanced" (grown by splitting random leaves).	'binary'
branch_length	float	Branch length. Ignored by "random_branches".	1.0
noise_sd	float	Error-bar scale relative to the spread of the data.	0.1
noise_spread	float	Each error bar is noise_sd times a log-uniform draw on [1 / noise_spread, noise_spread]. 1.0 is homoscedastic.	2.0
seed	int	Seed.	0

Returns:

Type	Description
SimulatedData	The tree and the data drawn on it.

simulate_counts

simulate_counts(
    n_leaves: int = 64,
    n_genes: int = 300,
    *,
    library_size: float = 3000.0,
    seed: int = 0,
) -> SimulatedCounts

UMI counts drawn on a known tree.

The noise-free leaf positions of simulate become log fold changes about a per-gene mean quotient, and counts are Poisson on top: count ~ Poisson(N_c * exp(log_q_g + x_gc)). Library sizes N_c are log-normal about library_size.

Parameters:

Name	Type	Description	Default
n_leaves	int	Cells, a power of two.	64
n_genes	int	Genes.	300
library_size	float	Median UMIs per cell.	3000.0
seed	int	Seed for both the tree and the counts.	0

Returns:

Type	Description
SimulatedCounts	The tree, the counts and the library sizes they were drawn with.