Lambda indexer for anndata

[grst]

With pandas data frames a pattern I use very commonly is

df_with_very_long_descriptive_name.loc[lambda x: x["fruit"] == "banana", :]

I was wondering if it would be possible to have lambda-based indexers for anndata as well, e.g.

adata_t = adata[lambda x: x.obs["cell_type"] == "T cell", :]

The benefits of this approach are imo:

• no need for duplication of the object name (which becomes annoying if it is not just named adata)
• reduction of copy&paste errors. I somewhat often end up with something like adata_t[adata.obs["something"], :], because I forget to update the second variable name.

[dawe]

Why not simply

adata_t = adata[adata.obs.query("cell_type == ‘T cell’").index]

?

[grst]

Because i need to type adata twice in the same expression ;)

On Wed, Apr 13, 2022, 20:16 Davide Cittaro @.***> wrote:

Why not simply

adata_t = adata[adata.obs.query("cell_type == ‘T cell’").index]

?

— Reply to this email directly, view it on GitHub https://github.com/scverse/anndata/issues/758#issuecomment-1098346546, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABVZRV4LGBP3OQCRJ4JPMDTVE4FQTANCNFSM5TLNZS7Q . You are receiving this because you authored the thread.Message ID: @.***>

[flying-sheep]

Great idea, pipelining does need to be possible.

[ivirshup]

I think this has been discussed before, and would definltley be in favor of something in this direction. I had thought maybe via a .select method.

I'm not sure I love lambda as the recommended way to do something like this though. It's only situationally more concise than typing adata twice, and I think we could ask get more out of alternative approaches.

I've been wondering about doing something more like polars or datafusion. These would go great with path based access too.

[flying-sheep]

It’s not verbosity, it’s about reusing an object in a pipeline without interrupting a pipleline and writing imperative code.

But yes, declarative code like polars is even better!

[ivirshup]

Rough sketch of a potential API:

def select(
    self, 
    identifiers: Union[str, list[str]] = "*",
    *,
    obs: Optional[Idxer] = None,
    var: Optional[Idxer] = None,
    copy: bool = False,
):
    """
    Return a new AnnData with selected elements at selected indices.

    By default, does not copy data unless necessary.

    Usage
    -----

    >>> adata.select(
            ["obsm/X_pca", "obs/cell_type"],
            obs=po.col("cell_type") == "B Cell",
        )
    AnnData object with n_obs × n_vars = 342 × 13714
        obs: "cell_type”
        obsm: "X_pca"
    """
    ...