R and Python side by side • DenMark

The R package DenMark and the Python package denmark build the same Stan data and fit the same Stan models, and they share their function names where Python allows it. The table maps each task to its function in both languages; the notes below it list the remaining differences. The R functions are documented in the reference, and the Python functions in their docstrings (help(dm.DenMark)).

Functions

Task	R	Python
Cells from an analysis object	`denmark_from_seurat(object, celltype)` (Seurat)	`from_anndata(adata, celltype_key)` (AnnData, SpatialData tables)
Gene filter	`denmark_filter_genes(expr)`	`filter_genes(expr)`
Tissue boundary	`denmark_alpha_shape(x, y)`	`alpha_shape(x, y)`
Grid	`buildgridpp(grid_size, dataset, gene)`	`buildgridpp(grid_size, dataset, gene)`
Grid centroids	`grid_pts_centroid(grid_size, dataset)`	`grid_pts_centroid(grid_size, dataset)`
HSGP basis indices	`denmark_hsgp_indices(basis)`	`hsgp_indices(basis)`
Stan data	`denmark_prepare()`	`prepare_data()`
Compile a model	`denmark_model()`	`compile_model()`
Initial values	`denmark_init()`, which returns a function	`make_init()`, which returns a dict
Fit one gene	`DenMark()`, `DenMarkmain()`	`DenMark()` or `denmark()`, `DenMarkmain()`
Posterior summary	`denmark_summary(fit)`	`fit.summary()`
Harmonized draws	`denmark_draws(fit)`	`fit.draws()`, `fit.rho`
Latent fields	`denmark_fields(fit)`	`fit.fields()`
Correlation decay	`denmark_decay(fit)`	`fit.decay()`
HSGP check	`denmark_check_hsgp(fit)`	`fit.check_hsgp()`
Gene scan	`denmark_scan()`, `workers =` for parallel genes	`scan_genes()`, `n_jobs=` for parallel genes
Results into the analysis object	`denmark_annotate(object, scan)`	`annotate(adata, scan)`
Save and reload	`denmark_save()`, `denmark_load()` (RDS)	`save_result()`, `load_result()` (pickle)
Simulation	`denmark_simulate()`	`simulate()`
Toy data	`denmark_toy_data()`	`denmark.datasets.load_toy()`
Figures	`plot_denmark_boundary()`, `plot_denmark_grid()`, `plot_denmark_fields()`, `plot_denmark_decay()`, `plot_denmark_scan()` (ggplot2)	`plot_boundary()`, `plot_grid()`, `plot_fields()`, `plot_decay()`, `plot_scan()` (matplotlib figures)
Matérn-3/2 covariance	`matern32(d, sigma, rho)`	`matern32(d, sigma, ell)`

Differences to keep in mind

A fit returns a DenMarkResult, whose methods take the place of the R post-processing functions.
Grid indices start at 1 in R (grid_index = i + (j - 1) * nx) and at 0 in Python (grid_index = i + j * nx). In both, x varies fastest.
The tissue boundary is an sf polygon with the attribute alpha in R and a TissueBoundary in Python. Both languages take it as boundary in the same functions.
A few arguments have different names: DenMark() takes inits in Python and init in R, and compile_model() takes force_compile where denmark_model() takes force_recompile. The plot functions take xlabel and ylabel where R takes xlab and ylab, and point_size in plot_boundary() is a matplotlib marker area in points squared rather than a ggplot2 point size.
With a tissue boundary, the grid areas agree between the languages up to the last bit: sf computes polygon areas with GDAL’s formula and shapely with GEOS’s. Grid indices, types, counts and coordinates are identical.
The third argument of matern32() is the length-scale in both languages. R calls it rho for backward compatibility; it is not the density-expression correlation.
simulate() draws with numpy’s generator, so it does not reproduce denmark_simulate() for the same seed. To compare the two languages on the same data, read the same file, as the toy data do.
With save_dir, scan_genes() caches one CSV row per gene and denmark_scan() one RDS file per gene.
keep_fit=False keeps only the draws of the key parameters, which saves memory in large scans; fit.fields() then needs a refit.
DenMark(stan_file=...) fits a custom Stan file, which must declare the same data and parameters as DenMark.stan, or as DenMark-vararea.stan when a boundary is given.