multidms fitting bug in func_effect_shifts.ipynb

[jgallowa07]

While working on #86 , we encountered a bug in the lasso sweep code in the notebook

lasso_shifts = params["lasso_shifts"]
assert len(lasso_shifts) == len(set(lasso_shifts))
n_lasso = len(params["lasso_shifts"])
fig, ax = plt.subplots(n_lasso, 2, figsize=[6, n_lasso * 3])

model = multidms.Model(data)
mutations_df = []
for i, lasso_shift in enumerate(lasso_shifts):
    lasso_shift = float(lasso_shift)
    print(f"Fitting model for {lasso_shift=}")
    model.fit(lasso_shift=lasso_shift)
    mutations_df.append(
        model.get_mutations_df(phenotype_as_effect=True).assign(lasso_shift=lasso_shift)
    )
    model.plot_epistasis(ax=ax[i, 1], alpha=0.1, show=False, legend=not i)
    model.plot_pred_accuracy(ax=ax[i, 0], alpha=0.1, show=False, legend=False)
    ax[i, 1].set_title(f"Epistasis fit (lasso {lasso_shift})")
    ax[i, 0].set_title(f"Accuracy (lasso {lasso_shift})")

plt.show()

Here, the model object is only initialized once, meaning there is only a single set of parameters being re-fit for each of the lasso strengths in the loop. To fix this, we should simply need to move the model instantiation inside the loop.

[jbloom]

I merged the pull request that fixes this. But I would note, although I agree this new code makes sense, I'm not sure the prior code was truly a bug, was it? Starting each lasso from the results of the prior weight is also an arguably valid approach even if we think fitting each one independently is better, right? Or is there a reason why it was strictly speaking a bug before?

[Caleb-Carr]

@jbloom What I noticed is that the results varied pretty substantially depending on what values were used for the lasso sweep even though the 'final' analysis used the same lasso weight. I attached an image that shows how different lasso sweeps affect the final analysis even though each one uses the same 'final' lasso weight. I think the main downside is that this could be pretty confusing if you expect similar results for a particular lasso weight but they differ a lot depending on the other weights tested?

[jbloom]

Anyway, should have behavior you want now.