Switch to formulaic-contrasts

[grst]

Hi @BorisMuzellec,

I created formulaic-contrasts containing all the code required for introspecting formulaic models and defining contrasts.

As described here it can be used either by inheritance or composition. Using inheritance would save us from redefining .variables, .cond and .contrast, but there'd be a name clash with the design variable so I went for the composition approach instead. Using inheritance/mixin would be possible with a little bit of refactoring, but I'd leave that to you should you prefer this approach.

There's now also documentation on how to build contrasts: https://formulaic-contrasts.readthedocs.io/en/latest/contrasts.html
Feel free to reference that from and/or copy it to the PyDESeq2 docs.

Best,
Gregor

[BorisMuzellec]

Hi @grst,

Thanks for creating this PR! I'm happy with the way you propose to reorganize things :)

I'll add a reference to https://formulaic-contrasts.readthedocs.io/en/latest/contrasts.html in the docs when I'll have some time, let me open an issue to remember to do it

[grst]

hang on, is this what you expect? it should be [variable, baseline, group_to_compare], right? I just want to make sure I didn't mix up things in formulaic-contrasts

[BorisMuzellec]

It should be [variable, group_to_compare, baseline], and I think that you are right and that I got confused in #328

It used to be optional to input a contrast to DeseqStats before #328. I went a bit too fast in #328 and put the contrasts that passed the tests, but I realised that the code used to be

def _contrast(self, column: str, baseline: str, group_to_compare: str) -> np.ndarray:
    """Build a simple contrast for pairwise comparisons.
    This is equivalent to
    ```
    model.cond(<column> = baseline) - model.cond(<column> = group_to_compare)
    ```
    Parameters
    ----------
    column: str
        The column to contrast.
    baseline: str
        The baseline group.
    group_to_compare: str
        The group to compare to the baseline.
    Returns
    -------
    np.ndarray
        The contrast vector.
    """
    return self.dds.cond(**{column: baseline}) - self.dds.cond(
        **{column: group_to_compare}
    )

Which (unless I'm mistaken again!) was wrong (it should be
self.dds.cond(**{column: group_to_compare}) - self.dds.cond(**{column: baseline}))

[BorisMuzellec]

Which is indeed what is implemented in pertpy: https://github.com/scverse/pertpy/blob/0091d0dcb6139449a5a67a1255635afd0e9e4cfb/pertpy/tools/_differential_gene_expression/_base.py#L1089