structboost.plot_dimension_correlation

structboost.plot_dimension_correlation(adata, *, dims=None, method='pearson', absolute=False, latent_key='X_bae', annotate=None, cmap=None, figsize=None, dpi=150)[source]

Heatmap of the correlation between latent dimensions.

Whether two dimensions are near-duplicates decides whether they can be read as two findings or one, and nothing else in the readout answers it – every other panel looks at one dimension at a time.

It matters most when the fit does not enforce separation: BAEConfig defaults disentanglement to "orthogonal", and a fit that turns it off can carry two dimensions describing the same programme with nothing else flagging it.

Parameters:

  • adata – AnnData with a fitted BAE, or an (n_cells, n_dims) array of scores.

  • dims (Sequence[int] | None) – Dimensions to include. None uses all.

  • method (Literal['pearson', 'spearman']) – "pearson" (default) or "spearman". Spearman is computed as Pearson on tie-averaged ranks, which matters here: a sparse encoder leaves many cells at exactly zero on a subgroup dimension.

  • absolute (bool) – Plot |r| instead of r. Signed values get a diverging map centred on zero, because the sign of a correlation is meaningful and a latent dimension’s own sign is arbitrary; absolute values are a magnitude and get a sequential one.

  • annotate (bool | None) – Write each value into its cell. None annotates when there are at most 12 dimensions, beyond which the numbers stop fitting.

  • cmap – Override the colour map chosen by absolute.

  • figsize (tuple[float, float] | None) – Overrides for the computed size, and the figure’s resolution.

  • dpi (int) – Overrides for the computed size, and the figure’s resolution.

  • latent_key (str)

Returns:

fig, ax

Raises:

  • KeyError – If latent_key is missing from an AnnData input.

  • ValueError – If dims is out of range or method is unknown.

Examples

>>> plot_dimension_correlation(adata, method="spearman")
>>> plot_dimension_correlation(adata, absolute=True)