Align Mixscape with Seurat’s implementation

[Lilly-May]

PR Checklist

• Referenced issue is linked
• If you've fixed a bug or added code that should be tested, add tests!
• Documentation in docs is updated

Description of changes

I made the following updates to pt.tl.Mixscape():

• Added a de_layer parameter to mixscape(), since DEG computation should be based on adata.X, while the rest of the method operates on adata.layers[X_pert], i.e., the perturbation signature. Seurat’s implementation also includes this parameter (see here).
• Added a test_method parameter to mixscape() and lda() to specify the test used for DEG computation. Seurat uses Wilcoxon by default (see here), while pertpy previously always used a t-test. Hence, users can now choose their preferred method.
• Added a scale parameter to mixscape. By default, Seurat scales DEG expression within the respective group (see here), so I introduced an option to enable this in pertpy too, which is set to True by default.
• Fixed an issue in the loop that assigns cells to NP and KO. Previously, the loop always used the original labels at the beginning of each iteration instead of updating them based on the previous iteration’s results. Now, it correctly updates the labels until convergence. This was also mentioned in issue Mixscape classification #688.
• Implemented a CustomGaussianMixture model. mixscape() fits a Gaussian Mixture Model to perturbed and non-perturbed cells, which is then used to assign cells to NP or KO. However, Seurat’s model fixes the mean and standard deviation for the NT distributions (see here), whereas Scikit-learn’s GaussianMixture does not support this. Hence, so far, our implementation effectively fit two distributions instead of only one as in Seurat's case. To address this, I created a CustomGaussianMixture class that inherits from GaussianMixture and overrides the M-step of the EM algorithm, allowing to fix certain mean and/or covariance values.
• Updated Gaussian Mixture Model initialization to align with Seurat’s approach. Seurat’s model allows specifying initial standard deviation values, while Scikit-learn’s implementation specifies precision (inverse of variance). I adjusted our initialization so that we now have the same behavior as in Seurat.

[Zethson]

Many great improvements! Thank you so much