Default parameters for generation of synthetic DIALOGUE data
params_sc_synthetic_dialogue.RdShapes the fixture with a planted multicellular programme that
generate_dialogue_test_data() builds. The defaults give 14
samples x 25 cells x 3 cell types = 1050 cells over 400 genes.
Usage
params_sc_synthetic_dialogue(
n_samples = 14L,
cells_per_sample = 25L,
n_cell_types = 3L,
n_features = 8L,
n_sample_features = 5L,
n_genes = 400L,
n_planted = 8L
)Arguments
- n_samples
Integer. Samples the experiment spans. DIALOGUE needs at least 5. Defaults to
14L.- cells_per_sample
Integer. Cells per sample per cell type. Defaults to
25L.- n_cell_types
Integer. Number of cell types. Must be at least 2. Defaults to
3L.- n_features
Integer. Feature columns per cell type. Must be at least 2. Defaults to
8L.- n_sample_features
Integer. Feature columns carrying a per-sample component. The first of those is the shared programme, the rest are cell-type-specific nuisance; anything past this count is pure noise and exists so the ANOVA filter has something to reject. Defaults to
5L.- n_genes
Integer. Number of genes. Defaults to
400L.- n_planted
Integer. Planted genes per cell type. The blocks are contiguous, so
n_planted * n_cell_typeshas to fit inton_genes. Defaults to8L.
Value
A named list with the following elements:
n_samples - Integer. Samples the experiment spans. DIALOGUE needs at least 5. Defaults to
14L.cells_per_sample - Integer. Cells per sample per cell type. Defaults to
25L.n_cell_types - Integer. Number of cell types. Must be at least 2. Defaults to
3L.n_features - Integer. Feature columns per cell type. Must be at least 2. Defaults to
8L.n_sample_features - Integer. Feature columns carrying a per-sample component. The first of those is the shared programme, the rest are cell-type-specific nuisance; anything past this count is pure noise and exists so the ANOVA filter has something to reject. Defaults to
5L.n_genes - Integer. Number of genes. Defaults to
400L.n_planted - Integer. Planted genes per cell type. The blocks are contiguous, so
n_planted * n_cell_typeshas to fit inton_genes. Defaults to8L.
Details
n_genes defaults higher here than on the Rust side, which plants the same
structure into 90. R runs the counts through the normal ingestion path, and
on a small panel the planted genes are a large enough share of the library
that log-normalisation divides the signal back out: at 90 genes the library
size tracks the programme at an r of 0.75 and background genes pick up a
spurious correlation of their own. At 400 the planted block is a few percent
of the library and the contrast survives. The Rust tests feed the planted
layer straight in, so they never meet this.