Please describe your wishes and possible alternatives to achieve the desired result.

Hi,

I would like to propose adding built-in support for custom accessors in AnnData. This feature would allow users to register and use dedicated namespaces on AnnData objects without the need for subclassing or mixins, similar to the mechanisms available in Polars and Xarray.

Motivation

Custom accessors provide a clean and modular way to extend the functionality of a core class. For example, in both Polars and Xarray, users can define custom functionality and then access it through a dedicated namespace (e.g., df.my_accessor.my_method()).

This enables:

• Enhanced Modularity: Custom functionality can be isolated in separate modules.
• Ease of Use: Users extend AnnData objects with additional methods without modifying or subclassing the core class.
• Extensibility: Third-party developers can integrate domain-specific functionality that naturally augments the core AnnData behavior.

Example

import anndata as ad


@ad.extensions.register_anndata_namespace("transforms")
class Transforms:
    def __init__(self, adata: ad.AnnData):
        self._adata = adata

    def arcsinh(
        self, shift: float, scale: float, layer: str, inplace: bool = False
    ) -> ad.AnnData | None:
        """Apply an arcsinh transformation to the X matrix."""
        import numpy as np

        # Example transformation logic
        transformed = np.arcsinh(self._adata.X / scale) + shift
        return None if inplace else self._adata

and you'd use it like so:

adata = ad.AnnData(X=...)
adata.transforms.arcsinh(shift=1, scale=1, layer='arcsinh', inplace=True)

Existing examples out in the scverse

Similar accessor functionality has been implemented in other scverse projects:

• Spatialdata-Plot implements accessors using Xarray’s approach, providing a nice to use API for plotting functionality.

• holoviz-topics/hv-anndata uses holoviz's interface solution to extend AnnData.