crested.pl.corr.violin

crested.pl.corr.violin(adata, model_names=None, split='test', log_transform=False, plot_kws=None, ax=None, **kwargs)

Plot correlation violinplots of predictions vs ground truth for different cell types.

Calculates correlations over cell types, i.e. for each cell type, calculates the correlation between ground truth and predictions across all regions.

Parameters:

• adata (AnnData) – AnnData object containing the data in X and predictions in layers.

• model_names (str | list[str] | None (default: None)) – Model name or list of model names (adata.layers) to use to calculate correlations with ground truths. Default is to plot all models in adata.layers.

• split (str | None (default: 'test')) – ‘train’, ‘val’, ‘test’ subset or None. If None, will use all targets. If not None, expects a “split” column in adata.var.

• log_transform (bool (default: False)) – Whether to log-transform the data before calculating correlations.

• plot_kws (dict | None (default: None)) – Extra keyword arguments passed to violinplot(). Defaults: {'inner': 'point', 'orient': 'v'}.

• ax (Axes | None (default: None)) – Axis to plot values on. If not supplied, creates a figure from scratch.

• width – Width of the newly created figure if ax=None. Default width is 6, +1 for every model > 5.

• height – Height of the newly created figure if ax=None. Default is 8.

• kwargs – Additional arguments passed to render_plot() to control the final plot output. Please see render_plot() for details. Custom defaults for correlations: ylabel='Pearson correlation', grid='y', ylim=(0.0, 1.0), title="Class-wise prediction vs ground truth correlations", xtick_rotation=55 (if any model_names > 15 characters).

Return type:

tuple[Figure, Axes] | None

See also

crested.pl.render_plot, crested.pl.corr.heatmap

Examples

>>> crested.pl.corr.violin(
...     adata_specific,
...     model_names=["Base model", "Fine-tuned"],
...     split="test",
...     log_transform=True,
...     plot_kws={'palette': 'Paired', 'order': ["Base model", "Fine-tuned"]},
...     title="Per-class model correlation",
... )