pertpy 0.9.5__py3-none-any.whl → 0.11.0__py3-none-any.whl

pertpy/tools/_mixscape.py

@@ -9,15 +9,14 @@ import numpy as np
 import pandas as pd
 import scanpy as sc
 import seaborn as sns
+from fast_array_utils.stats import mean, mean_var
 from scanpy import get
-from scanpy._settings import settings
 from scanpy._utils import _check_use_raw, sanitize_anndata
 from scanpy.plotting import _utils
 from scanpy.tools._utils import _choose_representation
-from scipy.sparse import csr_matrix, issparse, spmatrix
+from scipy.sparse import csr_matrix, spmatrix
 from sklearn.mixture import GaussianMixture
 
-import pertpy as pt
 from pertpy._doc import _doc_params, doc_common_plot_args
 
 if TYPE_CHECKING:
@@ -41,9 +40,12 @@ class Mixscape:
         adata: AnnData,
         pert_key: str,
         control: str,
+        *,
+        ref_selection_mode: Literal["nn", "split_by"] = "nn",
         split_by: str | None = None,
         n_neighbors: int = 20,
         use_rep: str | None = None,
+        n_dims: int | None = 15,
         n_pcs: int | None = None,
         batch_size: int | None = None,
         copy: bool = False,
@@ -51,14 +53,18 @@ class Mixscape:
     ):
         """Calculate perturbation signature.
 
-        For each cell, we identify `n_neighbors` cells from the control pool with the most similar mRNA expression profiles.
-        The perturbation signature is calculated by subtracting the averaged mRNA expression profile of the control
-        neighbors from the mRNA expression profile of each cell.
+        The perturbation signature is calculated by subtracting the mRNA expression profile of each cell from the averaged
+        mRNA expression profile of the control cells (selected according to `ref_selection_mode`).
+        The implementation resembles https://satijalab.org/seurat/reference/runmixscape. Note that in the original implementation, the
+        perturbation signature is calculated on unscaled data by default, and we therefore recommend to do the same.
 
         Args:
             adata: The annotated data object.
             pert_key: The column  of `.obs` with perturbation categories, should also contain `control`.
-            control: Control category from the `pert_key` column.
+            control: Name of the control category from the `pert_key` column.
+            ref_selection_mode: Method to select reference cells for the perturbation signature calculation. If `nn`,
+                the `n_neighbors` cells from the control pool with the most similar mRNA expression profiles are selected. If `split_by`,
+                the control cells from the same split in `split_by` (e.g. indicating biological replicates) are used to calculate the perturbation signature.
             split_by: Provide the column `.obs` if multiple biological replicates exist to calculate
                 the perturbation signature for every replicate separately.
             n_neighbors: Number of neighbors from the control to use for the perturbation signature.
@@ -66,7 +72,10 @@ class Mixscape:
                 If `None`, the representation is chosen automatically:
                 For `.n_vars` < 50, `.X` is used, otherwise 'X_pca' is used.
                 If 'X_pca' is not present, it’s computed with default parameters.
-            n_pcs: Use this many PCs. If `n_pcs==0` use `.X` if `use_rep is None`.
+            n_dims: Number of dimensions to use from the representation to calculate the perturbation signature.
+                If `None`, use all dimensions.
+            n_pcs: If PCA representation is used, the number of principal components to compute.
+                If `n_pcs==0` use `.X` if `use_rep is None`.
             batch_size: Size of batch to calculate the perturbation signature.
                 If 'None', the perturbation signature is calcuated in the full mode, requiring more memory.
                 The batched mode is very inefficient for sparse data.
@@ -83,8 +92,13 @@ class Mixscape:
             >>> import pertpy as pt
             >>> mdata = pt.dt.papalexi_2021()
             >>> ms_pt = pt.tl.Mixscape()
-            >>> ms_pt.perturbation_signature(mdata["rna"], "perturbation", "NT", "replicate")
+            >>> ms_pt.perturbation_signature(mdata["rna"], "perturbation", "NT", split_by="replicate")
         """
+        if ref_selection_mode not in ["nn", "split_by"]:
+            raise ValueError("ref_selection_mode must be either 'nn' or 'split_by'.")
+        if ref_selection_mode == "split_by" and split_by is None:
+            raise ValueError("split_by must be provided if ref_selection_mode is 'split_by'.")
+
         if copy:
             adata = adata.copy()
 
@@ -92,59 +106,73 @@ class Mixscape:
 
         control_mask = adata.obs[pert_key] == control
 
-        if split_by is None:
-            split_masks = [np.full(adata.n_obs, True, dtype=bool)]
+        if ref_selection_mode == "split_by":
+            for split in adata.obs[split_by].unique():
+                split_mask = adata.obs[split_by] == split
+                control_mask_group = control_mask & split_mask
+                control_mean_expr = mean(adata.X[control_mask_group], axis=0)
+                adata.layers["X_pert"][split_mask] = (
+                    np.repeat(control_mean_expr.reshape(1, -1), split_mask.sum(), axis=0)
+                    - adata.layers["X_pert"][split_mask]
+                )
         else:
-            split_obs = adata.obs[split_by]
-            split_masks = [split_obs == cat for cat in split_obs.unique()]
+            if split_by is None:
+                split_masks = [np.full(adata.n_obs, True, dtype=bool)]
+            else:
+                split_obs = adata.obs[split_by]
+                split_masks = [split_obs == cat for cat in split_obs.unique()]
 
-        representation = _choose_representation(adata, use_rep=use_rep, n_pcs=n_pcs)
+            representation = _choose_representation(adata, use_rep=use_rep, n_pcs=n_pcs)
+            if n_dims is not None and n_dims < representation.shape[1]:
+                representation = representation[:, :n_dims]
 
-        for split_mask in split_masks:
-            control_mask_split = control_mask & split_mask
+            from pynndescent import NNDescent
 
-            R_split = representation[split_mask]
-            R_control = representation[np.asarray(control_mask_split)]
+            for split_mask in split_masks:
+                control_mask_split = control_mask & split_mask
 
-            from pynndescent import NNDescent
+                R_split = representation[split_mask]
+                R_control = representation[np.asarray(control_mask_split)]
 
-            eps = kwargs.pop("epsilon", 0.1)
-            nn_index = NNDescent(R_control, **kwargs)
-            indices, _ = nn_index.query(R_split, k=n_neighbors, epsilon=eps)
+                eps = kwargs.pop("epsilon", 0.1)
+                nn_index = NNDescent(R_control, **kwargs)
+                indices, _ = nn_index.query(R_split, k=n_neighbors, epsilon=eps)
 
-            X_control = np.expm1(adata.X[np.asarray(control_mask_split)])
+                X_control = np.expm1(adata.X[np.asarray(control_mask_split)])
 
-            n_split = split_mask.sum()
-            n_control = X_control.shape[0]
+                n_split = split_mask.sum()
+                n_control = X_control.shape[0]
 
-            if batch_size is None:
-                col_indices = np.ravel(indices)
-                row_indices = np.repeat(np.arange(n_split), n_neighbors)
+                if batch_size is None:
+                    col_indices = np.ravel(indices)
+                    row_indices = np.repeat(np.arange(n_split), n_neighbors)
 
-                neigh_matrix = csr_matrix(
-                    (np.ones_like(col_indices, dtype=np.float64), (row_indices, col_indices)),
-                    shape=(n_split, n_control),
-                )
-                neigh_matrix /= n_neighbors
-                adata.layers["X_pert"][split_mask] -= np.log1p(neigh_matrix @ X_control)
-            else:
-                is_sparse = issparse(X_control)
-                split_indices = np.where(split_mask)[0]
-                for i in range(0, n_split, batch_size):
-                    size = min(i + batch_size, n_split)
-                    select = slice(i, size)
+                    neigh_matrix = csr_matrix(
+                        (np.ones_like(col_indices, dtype=np.float64), (row_indices, col_indices)),
+                        shape=(n_split, n_control),
+                    )
+                    neigh_matrix /= n_neighbors
+                    adata.layers["X_pert"][np.asarray(split_mask)] = (
+                        sc.pp.log1p(neigh_matrix @ X_control) - adata.layers["X_pert"][np.asarray(split_mask)]
+                    )
+                else:
+                    split_indices = np.where(split_mask)[0]
+                    for i in range(0, n_split, batch_size):
+                        size = min(i + batch_size, n_split)
+                        select = slice(i, size)
 
-                    batch = np.ravel(indices[select])
-                    split_batch = split_indices[select]
+                        batch = np.ravel(indices[select])
+                        split_batch = split_indices[select]
 
-                    size = size - i
+                        size = size - i
 
-                    # sparse is very slow
-                    means_batch = X_control[batch]
-                    means_batch = means_batch.toarray() if is_sparse else means_batch
-                    means_batch = means_batch.reshape(size, n_neighbors, -1).mean(1)
+                        means_batch = X_control[batch]
+                        batch_reshaped = means_batch.reshape(size, n_neighbors, -1)
+                        means_batch, _ = mean_var(batch_reshaped, axis=1)
 
-                    adata.layers["X_pert"][split_batch] -= np.log1p(means_batch)
+                        adata.layers["X_pert"][split_batch] = (
+                            np.log1p(means_batch) - adata.layers["X_pert"][split_batch]
+                        )
 
         if copy:
             return adata
@@ -154,34 +182,44 @@ class Mixscape:
         adata: AnnData,
         labels: str,
         control: str,
+        *,
         new_class_name: str | None = "mixscape_class",
-        min_de_genes: int | None = 5,
         layer: str | None = None,
+        min_de_genes: int | None = 5,
         logfc_threshold: float | None = 0.25,
+        de_layer: str | None = None,
+        test_method: str | None = "wilcoxon",
         iter_num: int | None = 10,
+        scale: bool | None = True,
         split_by: str | None = None,
         pval_cutoff: float | None = 5e-2,
         perturbation_type: str | None = "KO",
+        random_state: int | None = 0,
         copy: bool | None = False,
+        **gmmkwargs,
     ):
         """Identify perturbed and non-perturbed gRNA expressing cells that accounts for multiple treatments/conditions/chemical perturbations.
 
-        The implementation resembles https://satijalab.org/seurat/reference/runmixscape
+        The implementation resembles https://satijalab.org/seurat/reference/runmixscape.
 
         Args:
             adata: The annotated data object.
             labels: The column of `.obs` with target gene labels.
-            control: Control category from the `pert_key` column.
+            control: Control category from the `labels` column.
             new_class_name: Name of mixscape classification to be stored in `.obs`.
-            min_de_genes: Required number of genes that are differentially expressed for method to separate perturbed and non-perturbed cells.
             layer: Key from adata.layers whose value will be used to perform tests on. Default is using `.layers["X_pert"]`.
+            min_de_genes: Required number of genes that are differentially expressed for method to separate perturbed and non-perturbed cells.
             logfc_threshold: Limit testing to genes which show, on average, at least X-fold difference (log-scale) between the two groups of cells (default: 0.25).
+            de_layer: Layer to use for identifying differentially expressed genes. If `None`, adata.X is used.
+            test_method: Method to use for differential expression testing.
             iter_num: Number of normalmixEM iterations to run if convergence does not occur.
-            split_by: Provide the column `.obs` if multiple biological replicates exist to calculate
-                    the perturbation signature for every replicate separately.
+            scale: Scale the data specified in `layer` before running the GaussianMixture model on it.
+            split_by: Provide `.obs` column with experimental condition/cell type annotation, if perturbations are condition/cell type-specific.
             pval_cutoff: P-value cut-off for selection of significantly DE genes.
             perturbation_type: specify type of CRISPR perturbation expected for labeling mixscape classifications.
+            random_state: Random seed for the GaussianMixture model.
             copy: Determines whether a copy of the `adata` is returned.
+            **gmmkwargs: Passed to custom implementation of scikit-learn Gaussian Mixture Model.
 
         Returns:
             If `copy=True`, returns the copy of `adata` with the classification result in `.obs`.
@@ -203,8 +241,8 @@ class Mixscape:
             >>> import pertpy as pt
             >>> mdata = pt.dt.papalexi_2021()
             >>> ms_pt = pt.tl.Mixscape()
-            >>> ms_pt.perturbation_signature(mdata["rna"], "perturbation", "NT", "replicate")
-            >>> ms_pt.mixscape(adata=mdata["rna"], control="NT", labels="gene_target", layer="X_pert")
+            >>> ms_pt.perturbation_signature(mdata["rna"], "perturbation", "NT", split_by="replicate")
+            >>> ms_pt.mixscape(mdata["rna"], "gene_target", "NT", layer="X_pert")
         """
         if copy:
             adata = adata.copy()
@@ -218,7 +256,16 @@ class Mixscape:
             split_masks = [split_obs == category for category in categories]
 
         perturbation_markers = self._get_perturbation_markers(
-            adata, split_masks, categories, labels, control, layer, pval_cutoff, min_de_genes, logfc_threshold
+            adata=adata,
+            split_masks=split_masks,
+            categories=categories,
+            labels=labels,
+            control=control,
+            layer=de_layer,
+            pval_cutoff=pval_cutoff,
+            min_de_genes=min_de_genes,
+            logfc_threshold=logfc_threshold,
+            test_method=test_method,
         )
 
         adata_comp = adata
@@ -231,6 +278,7 @@ class Mixscape:
                 raise KeyError(
                     "No 'X_pert' found in .layers! Please run perturbation_signature first to calculate perturbation signature!"
                 ) from None
+
         # initialize return variables
         adata.obs[f"{new_class_name}_p_{perturbation_type.lower()}"] = 0
         adata.obs[new_class_name] = adata.obs[labels].astype(str)
@@ -241,10 +289,12 @@ class Mixscape:
             dtype=np.object_,
         )
         gv_list: dict[str, dict] = {}
+
+        adata.obs[f"{new_class_name}_p_{perturbation_type.lower()}"] = 0.0
         for split, split_mask in enumerate(split_masks):
             category = categories[split]
-            genes = list(set(adata[split_mask].obs[labels]).difference([control]))
-            for gene in genes:
+            gene_targets = list(set(adata[split_mask].obs[labels]).difference([control]))
+            for gene in gene_targets:
                 post_prob = 0
                 orig_guide_cells = (adata.obs[labels] == gene) & split_mask
                 orig_guide_cells_index = list(orig_guide_cells.index[orig_guide_cells])
@@ -253,63 +303,79 @@ class Mixscape:
 
                 if len(perturbation_markers[(category, gene)]) == 0:
                     adata.obs.loc[orig_guide_cells, new_class_name] = f"{gene} NP"
+
                 else:
                     de_genes = perturbation_markers[(category, gene)]
-                    de_genes_indices = self._get_column_indices(adata, list(de_genes))
+                    de_genes_indices = np.where(np.isin(adata.var_names, list(de_genes)))[0]
+
                     dat = X[np.asarray(all_cells)][:, de_genes_indices]
+                    if scale:
+                        dat = sc.pp.scale(dat)
+
                     converged = False
                     n_iter = 0
-                    old_classes = adata.obs[labels][all_cells]
+                    old_classes = adata.obs[new_class_name][all_cells]
+
+                    nt_cells_dat_idx = all_cells[all_cells].index.get_indexer(nt_cells[nt_cells].index)
+                    nt_cells_mean = np.mean(dat[nt_cells_dat_idx], axis=0)
+
                     while not converged and n_iter < iter_num:
                         # Get all cells in current split&Gene
-                        guide_cells = (adata.obs[labels] == gene) & split_mask
+                        guide_cells = (adata.obs[new_class_name] == gene) & split_mask
+
                         # get average value for each gene over all selected cells
                         # all cells in current split&Gene minus all NT cells in current split
                         # Each row is for each cell, each column is for each gene, get mean for each column
-                        vec = np.mean(X[np.asarray(guide_cells)][:, de_genes_indices], axis=0) - np.mean(
-                            X[np.asarray(nt_cells)][:, de_genes_indices], axis=0
-                        )
+                        guide_cells_dat_idx = all_cells[all_cells].index.get_indexer(guide_cells[guide_cells].index)
+                        guide_cells_mean = np.mean(dat[guide_cells_dat_idx], axis=0)
+                        vec = guide_cells_mean - nt_cells_mean
+
                         # project cells onto the perturbation vector
                         if isinstance(dat, spmatrix):
-                            pvec = np.sum(np.multiply(dat.toarray(), vec), axis=1) / np.sum(np.multiply(vec, vec))
+                            pvec = dat.dot(vec) / np.dot(vec, vec)
                         else:
-                            pvec = np.sum(np.multiply(dat, vec), axis=1) / np.sum(np.multiply(vec, vec))
+                            pvec = np.dot(dat, vec) / np.dot(vec, vec)
                         pvec = pd.Series(np.asarray(pvec).flatten(), index=list(all_cells.index[all_cells]))
+
                         if n_iter == 0:
                             gv = pd.DataFrame(columns=["pvec", labels])
                             gv["pvec"] = pvec
                             gv[labels] = control
                             gv.loc[guide_cells, labels] = gene
-                            if gene not in gv_list.keys():
+                            if gene not in gv_list:
                                 gv_list[gene] = {}
                             gv_list[gene][category] = gv
 
-                        guide_norm = self._define_normal_mixscape(pvec[guide_cells])
-                        nt_norm = self._define_normal_mixscape(pvec[nt_cells])
-                        means_init = np.array([[nt_norm[0]], [guide_norm[0]]])
-                        precisions_init = np.array([nt_norm[1], guide_norm[1]])
-                        mm = GaussianMixture(
+                        means_init = np.array([[pvec[nt_cells].mean()], [pvec[guide_cells].mean()]])
+                        std_init = np.array([pvec[nt_cells].std(), pvec[guide_cells].std()])
+                        mm = MixscapeGaussianMixture(
                             n_components=2,
                             covariance_type="spherical",
                             means_init=means_init,
-                            precisions_init=precisions_init,
+                            precisions_init=1 / (std_init**2),
+                            random_state=random_state,
+                            max_iter=100,
+                            fixed_means=[pvec[nt_cells].mean(), None],
+                            fixed_covariances=[pvec[nt_cells].std() ** 2, None],
+                            **gmmkwargs,
                         ).fit(np.asarray(pvec).reshape(-1, 1))
                         probabilities = mm.predict_proba(np.array(pvec[orig_guide_cells_index]).reshape(-1, 1))
                         lik_ratio = probabilities[:, 0] / probabilities[:, 1]
                         post_prob = 1 / (1 + lik_ratio)
+
                         # based on the posterior probability, assign cells to the two classes
-                        adata.obs.loc[
-                            [orig_guide_cells_index[cell] for cell in np.where(post_prob > 0.5)[0]], new_class_name
-                        ] = gene
-                        adata.obs.loc[
-                            [orig_guide_cells_index[cell] for cell in np.where(post_prob <= 0.5)[0]], new_class_name
-                        ] = f"{gene} NP"
+                        ko_mask = post_prob > 0.5
+                        adata.obs.loc[np.array(orig_guide_cells_index)[ko_mask], new_class_name] = gene
+                        adata.obs.loc[np.array(orig_guide_cells_index)[~ko_mask], new_class_name] = f"{gene} NP"
+
                         if sum(adata.obs[new_class_name][split_mask] == gene) < min_de_genes:
                             adata.obs.loc[guide_cells, new_class_name] = "NP"
                             converged = True
-                        if adata.obs[new_class_name][all_cells].equals(old_classes):
+                        current_classes = adata.obs[new_class_name][all_cells]
+                        if (current_classes == old_classes).all():
                             converged = True
-                        old_classes = adata.obs[new_class_name][all_cells]
+                        old_classes = current_classes
+
                         n_iter += 1
 
                     adata.obs.loc[(adata.obs[new_class_name] == gene) & split_mask, new_class_name] = (
@@ -317,9 +383,7 @@ class Mixscape:
                     )
 
                 adata.obs[f"{new_class_name}_global"] = [a.split(" ")[-1] for a in adata.obs[new_class_name]]
-                adata.obs.loc[orig_guide_cells_index, f"{new_class_name}_p_{perturbation_type.lower()}"] = np.round(
-                    post_prob
-                ).astype("int64")
+                adata.obs.loc[orig_guide_cells_index, f"{new_class_name}_p_{perturbation_type.lower()}"] = post_prob
         adata.uns["mixscape"] = gv_list
 
         if copy:
@@ -330,11 +394,13 @@ class Mixscape:
         adata: AnnData,
         labels: str,
         control: str,
+        *,
         mixscape_class_global: str | None = "mixscape_class_global",
         layer: str | None = None,
         n_comps: int | None = 10,
         min_de_genes: int | None = 5,
         logfc_threshold: float | None = 0.25,
+        test_method: str | None = "wilcoxon",
         split_by: str | None = None,
         pval_cutoff: float | None = 5e-2,
         perturbation_type: str | None = "KO",
@@ -347,12 +413,12 @@ class Mixscape:
             labels: The column of `.obs` with target gene labels.
             control: Control category from the `pert_key` column.
             mixscape_class_global: The column of `.obs` with mixscape global classification result (perturbed, NP or NT).
-            layer: Key from `adata.layers` whose value will be used to perform tests on.
-            control: Control category from the `pert_key` column.
+            layer: Layer to use for identifying differentially expressed genes. If `None`, adata.X is used.
             n_comps: Number of principal components to use.
             min_de_genes: Required number of genes that are differentially expressed for method to separate perturbed and non-perturbed cells.
             logfc_threshold: Limit testing to genes which show, on average, at least X-fold difference (log-scale) between the two groups of cells.
-            split_by: Provide the column `.obs` if multiple biological replicates exist to calculate
+            test_method: Method to use for differential expression testing.
+            split_by: Provide `.obs` column with experimental condition/cell type annotation, if perturbations are condition/cell type-specific.
             pval_cutoff: P-value cut-off for selection of significantly DE genes.
             perturbation_type: Specify type of CRISPR perturbation expected for labeling mixscape classifications.
             copy: Determines whether a copy of the `adata` is returned.
@@ -370,9 +436,9 @@ class Mixscape:
             >>> import pertpy as pt
             >>> mdata = pt.dt.papalexi_2021()
             >>> ms_pt = pt.tl.Mixscape()
-            >>> ms_pt.perturbation_signature(mdata["rna"], "perturbation", "NT", "replicate")
-            >>> ms_pt.mixscape(adata=mdata["rna"], control="NT", labels="gene_target", layer="X_pert")
-            >>> ms_pt.lda(adata=mdata["rna"], control="NT", labels="gene_target", layer="X_pert")
+            >>> ms_pt.perturbation_signature(mdata["rna"], "perturbation", "NT", split_by="replicate")
+            >>> ms_pt.mixscape(mdata["rna"], "gene_target", "NT", layer="X_pert")
+            >>> ms_pt.lda(mdata["rna"], "gene_target", "NT")
         """
         if copy:
             adata = adata.copy()
@@ -388,9 +454,8 @@ class Mixscape:
             categories = split_obs.unique()
             split_masks = [split_obs == category for category in categories]
 
-        mixscape_identifier = pt.tl.Mixscape()
         # determine gene sets across all splits/groups through differential gene expression
-        perturbation_markers = mixscape_identifier._get_perturbation_markers(
+        perturbation_markers = self._get_perturbation_markers(
             adata=adata,
             split_masks=split_masks,
             categories=categories,
@@ -400,10 +465,12 @@ class Mixscape:
             pval_cutoff=pval_cutoff,
             min_de_genes=min_de_genes,
             logfc_threshold=logfc_threshold,
+            test_method=test_method,
         )
         adata_subset = adata[
             (adata.obs[mixscape_class_global] == perturbation_type) | (adata.obs[mixscape_class_global] == control)
-        ].copy()
+        ]
+        X = adata_subset.X - adata_subset.X.mean(0)
         projected_pcs: dict[str, np.ndarray] = {}
         # performs PCA on each mixscape class separately and projects each subspace onto all cells in the data.
         for _, (key, value) in enumerate(perturbation_markers.items()):
@@ -415,16 +482,10 @@ class Mixscape:
                 ].copy()
                 sc.pp.scale(gene_subset)
                 sc.tl.pca(gene_subset, n_comps=n_comps)
-                sc.pp.neighbors(gene_subset)
-                # projects each subspace onto all cells in the data.
-                sc.tl.ingest(adata=adata_subset, adata_ref=gene_subset, embedding_method="pca")
-                projected_pcs[key[1]] = adata_subset.obsm["X_pca"]
+                # project cells into PCA space of gene_subset
+                projected_pcs[key[1]] = np.asarray(np.dot(X, gene_subset.varm["PCs"]))
         # concatenate all pcs into a single matrix.
-        for index, (_, value) in enumerate(projected_pcs.items()):
-            if index == 0:
-                projected_pcs_array = value
-            else:
-                projected_pcs_array = np.concatenate((projected_pcs_array, value), axis=1)
+        projected_pcs_array = np.concatenate(list(projected_pcs.values()), axis=1)
 
         clf = LinearDiscriminantAnalysis(n_components=len(np.unique(adata_subset.obs[labels])) - 1)
         clf.fit(projected_pcs_array, adata_subset.obs[labels])
@@ -445,12 +506,21 @@ class Mixscape:
         pval_cutoff: float,
         min_de_genes: float,
         logfc_threshold: float,
+        test_method: str,
     ) -> dict[tuple, np.ndarray]:
-        """Determine gene sets across all splits/groups through differential gene expression
+        """Determine gene sets across all splits/groups through differential gene expression.
 
         Args:
             adata: :class:`~anndata.AnnData` object
-            col_names: Column names to extract the indices for
+            split_masks: List of boolean masks for each split/group.
+            categories: List of split/group names.
+            labels: The column of `.obs` with target gene labels.
+            control: Control category from the `labels` column.
+            layer: Key from adata.layers whose value will be used to compare gene expression.
+            pval_cutoff: P-value cut-off for selection of significantly DE genes.
+            min_de_genes: Required number of genes that are differentially expressed for method to separate perturbed and non-perturbed cells.
+            logfc_threshold: Limit testing to genes which show, on average, at least X-fold difference (log-scale) between the two groups of cells.
+            test_method: Method to use for differential expression testing.
 
         Returns:
             Set of column indices.
@@ -459,21 +529,23 @@ class Mixscape:
         for split, split_mask in enumerate(split_masks):
             category = categories[split]
             # get gene sets for each split
-            genes = list(set(adata[split_mask].obs[labels]).difference([control]))
+            gene_targets = list(set(adata[split_mask].obs[labels]).difference([control]))
             adata_split = adata[split_mask].copy()
             # find top DE genes between cells with targeting and non-targeting gRNAs
             sc.tl.rank_genes_groups(
                 adata_split,
                 layer=layer,
                 groupby=labels,
-                groups=genes,
+                groups=gene_targets,
                 reference=control,
-                method="t-test",
+                method=test_method,
                 use_raw=False,
             )
-            # get DE genes for each gene
-            for gene in genes:
-                logfc_threshold_mask = adata_split.uns["rank_genes_groups"]["logfoldchanges"][gene] >= logfc_threshold
+            # get DE genes for each target gene
+            for gene in gene_targets:
+                logfc_threshold_mask = (
+                    np.abs(adata_split.uns["rank_genes_groups"]["logfoldchanges"][gene]) >= logfc_threshold
+                )
                 de_genes = adata_split.uns["rank_genes_groups"]["names"][gene][logfc_threshold_mask]
                 pvals_adj = adata_split.uns["rank_genes_groups"]["pvals_adj"][gene][logfc_threshold_mask]
                 de_genes = de_genes[pvals_adj < pval_cutoff]
@@ -483,33 +555,8 @@ class Mixscape:
 
         return perturbation_markers
 
-    def _get_column_indices(self, adata, col_names):
-        if isinstance(col_names, str):  # pragma: no cover
-            col_names = [col_names]
-
-        indices = []
-        for idx, col in enumerate(adata.var_names):
-            if col in col_names:
-                indices.append(idx)
-
-        return indices
-
-    def _define_normal_mixscape(self, X: np.ndarray | sparse.spmatrix | pd.DataFrame | None) -> list[float]:
-        """Calculates the mean and standard deviation of a matrix.
-
-        Args:
-            X: The matrix to calculate the properties for.
-
-        Returns:
-            Mean and standard deviation of the matrix.
-        """
-        mu = X.mean()
-        sd = X.std()
-
-        return [mu, sd]
-
     @_doc_params(common_plot_args=doc_common_plot_args)
-    def plot_barplot(  # pragma: no cover
+    def plot_barplot(  # pragma: no cover # noqa: D417
         self,
         adata: AnnData,
         guide_rna_column: str,
@@ -522,7 +569,6 @@ class Mixscape:
         legend_text_size: int = 8,
         legend_bbox_to_anchor: tuple[float, float] = None,
         figsize: tuple[float, float] = (25, 25),
-        show: bool = True,
         return_fig: bool = False,
     ) -> Figure | None:
         """Barplot to visualize perturbation scores calculated by the `mixscape` function.
@@ -548,8 +594,8 @@ class Mixscape:
             >>> import pertpy as pt
             >>> mdata = pt.dt.papalexi_2021()
             >>> ms_pt = pt.tl.Mixscape()
-            >>> ms_pt.perturbation_signature(mdata["rna"], "perturbation", "NT", "replicate")
-            >>> ms_pt.mixscape(adata=mdata["rna"], control="NT", labels="gene_target", layer="X_pert")
+            >>> ms_pt.perturbation_signature(mdata["rna"], "perturbation", "NT", split_by="replicate")
+            >>> ms_pt.mixscape(mdata["rna"], "gene_target", "NT", layer="X_pert")
             >>> ms_pt.plot_barplot(mdata["rna"], guide_rna_column="NT")
 
         Preview:
@@ -611,14 +657,13 @@ class Mixscape:
         fig.subplots_adjust(hspace=0.5, wspace=0.5)
         plt.tight_layout()
 
-        if show:
-            plt.show()
         if return_fig:
             return fig
+        plt.show()
         return None
 
     @_doc_params(common_plot_args=doc_common_plot_args)
-    def plot_heatmap(  # pragma: no cover
+    def plot_heatmap(  # pragma: no cover # noqa: D417
         self,
         adata: AnnData,
         labels: str,
@@ -630,7 +675,6 @@ class Mixscape:
         subsample_number: int | None = 900,
         vmin: float | None = -2,
         vmax: float | None = 2,
-        show: bool = True,
         return_fig: bool = False,
         **kwds,
     ) -> Figure | None:
@@ -656,8 +700,8 @@ class Mixscape:
             >>> import pertpy as pt
             >>> mdata = pt.dt.papalexi_2021()
             >>> ms_pt = pt.tl.Mixscape()
-            >>> ms_pt.perturbation_signature(mdata["rna"], "perturbation", "NT", "replicate")
-            >>> ms_pt.mixscape(adata=mdata["rna"], control="NT", labels="gene_target", layer="X_pert")
+            >>> ms_pt.perturbation_signature(mdata["rna"], "perturbation", "NT", split_by="replicate")
+            >>> ms_pt.mixscape(mdata["rna"], "gene_target", "NT", layer="X_pert")
             >>> ms_pt.plot_heatmap(
             ...     adata=mdata["rna"], labels="gene_target", target_gene="IFNGR2", layer="X_pert", control="NT"
             ... )
@@ -683,14 +727,13 @@ class Mixscape:
             **kwds,
         )
 
-        if show:
-            plt.show()
         if return_fig:
             return fig
+        plt.show()
         return None
 
     @_doc_params(common_plot_args=doc_common_plot_args)
-    def plot_perturbscore(  # pragma: no cover
+    def plot_perturbscore(  # pragma: no cover # noqa: D417
         self,
         adata: AnnData,
         labels: str,
@@ -702,7 +745,6 @@ class Mixscape:
         split_by: str = None,
         before_mixscape: bool = False,
         perturbation_type: str = "KO",
-        show: bool = True,
         return_fig: bool = False,
     ) -> Figure | None:
         """Density plots to visualize perturbation scores calculated by the `pt.tl.mixscape` function.
@@ -734,8 +776,8 @@ class Mixscape:
             >>> import pertpy as pt
             >>> mdata = pt.dt.papalexi_2021()
             >>> ms_pt = pt.tl.Mixscape()
-            >>> ms_pt.perturbation_signature(mdata["rna"], "perturbation", "NT", "replicate")
-            >>> ms_pt.mixscape(adata=mdata["rna"], control="NT", labels="gene_target", layer="X_pert")
+            >>> ms_pt.perturbation_signature(mdata["rna"], "perturbation", "NT", split_by="replicate")
+            >>> ms_pt.mixscape(mdata["rna"], "gene_target", "NT", layer="X_pert")
             >>> ms_pt.plot_perturbscore(adata=mdata["rna"], labels="gene_target", target_gene="IFNGR2", color="orange")
 
         Preview:
@@ -744,7 +786,7 @@ class Mixscape:
         if "mixscape" not in adata.uns:
             raise ValueError("Please run the `mixscape` function first.")
         perturbation_score = None
-        for key in adata.uns["mixscape"][target_gene].keys():
+        for key in adata.uns["mixscape"][target_gene]:
             perturbation_score_temp = adata.uns["mixscape"][target_gene][key]
             perturbation_score_temp["name"] = key
             if perturbation_score is None:
@@ -851,14 +893,13 @@ class Mixscape:
                 plt.legend(title="mixscape class", title_fontsize=14, fontsize=12)
                 sns.despine()
 
-        if show:
-            plt.show()
         if return_fig:
             return plt.gcf()
+        plt.show()
         return None
 
     @_doc_params(common_plot_args=doc_common_plot_args)
-    def plot_violin(  # pragma: no cover
+    def plot_violin(  # pragma: no cover # noqa: D417
         self,
         adata: AnnData,
         target_gene_idents: str | list[str],
@@ -879,7 +920,6 @@ class Mixscape:
         ylabel: str | Sequence[str] | None = None,
         rotation: float | None = None,
         ax: Axes | None = None,
-        show: bool = True,
         return_fig: bool = False,
         **kwargs,
     ) -> Axes | Figure | None:
@@ -910,8 +950,8 @@ class Mixscape:
             >>> import pertpy as pt
             >>> mdata = pt.dt.papalexi_2021()
             >>> ms_pt = pt.tl.Mixscape()
-            >>> ms_pt.perturbation_signature(mdata["rna"], "perturbation", "NT", "replicate")
-            >>> ms_pt.mixscape(adata=mdata["rna"], control="NT", labels="gene_target", layer="X_pert")
+            >>> ms_pt.perturbation_signature(mdata["rna"], "perturbation", "NT", split_by="replicate")
+            >>> ms_pt.mixscape(mdata["rna"], "gene_target", "NT", layer="X_pert")
             >>> ms_pt.plot_violin(
             ...     adata=mdata["rna"], target_gene_idents=["NT", "IFNGR2 NP", "IFNGR2 KO"], groupby="mixscape_class"
             ... )
@@ -939,7 +979,7 @@ class Mixscape:
             if len(ylabel) != 1:
                 raise ValueError(f"Expected number of y-labels to be `1`, found `{len(ylabel)}`.")
         elif len(ylabel) != len(keys):
-            raise ValueError(f"Expected number of y-labels to be `{len(keys)}`, " f"found `{len(ylabel)}`.")
+            raise ValueError(f"Expected number of y-labels to be `{len(keys)}`, found `{len(ylabel)}`.")
 
         if groupby is not None:
             if hue is not None:
@@ -992,7 +1032,7 @@ class Mixscape:
                 g.set(yscale="log")
             g.set_titles(col_template="{col_name}").set_xlabels("")
             if rotation is not None:
-                for ax in g.axes[0]:
+                for ax in g.axes[0]:  # noqa: PLR1704
                     ax.tick_params(axis="x", labelrotation=rotation)
         else:
             # set by default the violin plot cut=0 to limit the extend
@@ -1010,7 +1050,7 @@ class Mixscape:
             else:
                 axs = [ax]
             for ax, y, ylab in zip(axs, ys, ylabel, strict=False):
-                ax = sns.violinplot(
+                ax = sns.violinplot(  # noqa: PLW2901
                     x=x,
                     y=y,
                     data=obs_tidy,
@@ -1024,7 +1064,7 @@ class Mixscape:
                 # Get the handles and labels.
                 handles, labels = ax.get_legend_handles_labels()
                 if stripplot:
-                    ax = sns.stripplot(
+                    ax = sns.stripplot(  # noqa: PLW2901
                         x=x,
                         y=y,
                         data=obs_tidy,
@@ -1047,12 +1087,9 @@ class Mixscape:
                 if rotation is not None:
                     ax.tick_params(axis="x", labelrotation=rotation)
 
-        show = settings.autoshow if show is None else show
         if hue is not None and stripplot is True:
             plt.legend(handles, labels)
 
-        if show:
-            plt.show()
         if return_fig:
             if multi_panel and groupby is None and len(ys) == 1:
                 return g
@@ -1060,10 +1097,11 @@ class Mixscape:
                 return axs[0]
             else:
                 return axs
+        plt.show()
         return None
 
     @_doc_params(common_plot_args=doc_common_plot_args)
-    def plot_lda(  # pragma: no cover
+    def plot_lda(  # pragma: no cover # noqa: D417
         self,
         adata: AnnData,
         control: str,
@@ -1076,20 +1114,22 @@ class Mixscape:
         color_map: Colormap | str | None = None,
         palette: str | Sequence[str] | None = None,
         ax: Axes | None = None,
-        show: bool = True,
         return_fig: bool = False,
         **kwds,
     ) -> Figure | None:
         """Visualizing perturbation responses with Linear Discriminant Analysis. Requires `pt.tl.mixscape()` to be run first.
 
         Args:
-            adata: The annotated data object.
+            adata: The annotated data objectplot_heatmap.
             control: Control category from the `pert_key` column.
             mixscape_class: The column of `.obs` with the mixscape classification result.
             mixscape_class_global: The column of `.obs` with mixscape global classification result (perturbed, NP or NT).
             perturbation_type: Specify type of CRISPR perturbation expected for labeling mixscape classifications.
-            lda_key: If not specified, lda looks .uns["mixscape_lda"] for the LDA results.
             n_components: The number of dimensions of the embedding.
+            lda_key: If not specified, lda looks .uns["mixscape_lda"] for the LDA results.
+            color_map: Matplotlib color map.
+            palette: Matplotlib palette.
+            ax: Matplotlib axes.
             {common_plot_args}
             **kwds: Additional arguments to `scanpy.pl.umap`.
 
@@ -1097,9 +1137,9 @@ class Mixscape:
             >>> import pertpy as pt
             >>> mdata = pt.dt.papalexi_2021()
             >>> ms_pt = pt.tl.Mixscape()
-            >>> ms_pt.perturbation_signature(mdata["rna"], "perturbation", "NT", "replicate")
-            >>> ms_pt.mixscape(adata=mdata["rna"], control="NT", labels="gene_target", layer="X_pert")
-            >>> ms_pt.lda(adata=mdata["rna"], control="NT", labels="gene_target", layer="X_pert")
+            >>> ms_pt.perturbation_signature(mdata["rna"], "perturbation", "NT", split_by="replicate")
+            >>> ms_pt.mixscape(mdata["rna"], "gene_target", "NT", layer="X_pert")
+            >>> ms_pt.lda(mdata["rna"], "gene_target", "NT", split_by="replicate")
             >>> ms_pt.plot_lda(adata=mdata["rna"], control="NT")
 
         Preview:
@@ -1129,8 +1169,54 @@ class Mixscape:
             **kwds,
         )
 
-        if show:
-            plt.show()
         if return_fig:
             return fig
+        plt.show()
         return None
+
+
+class MixscapeGaussianMixture(GaussianMixture):
+    def __init__(
+        self,
+        n_components: int,
+        fixed_means: Sequence[float] | None = None,
+        fixed_covariances: Sequence[float] | None = None,
+        **kwargs,
+    ):
+        """Custom Gaussian Mixture Model where means and covariances can be fixed for specific components.
+
+        Args:
+            n_components: Number of Gaussian components
+            fixed_means: Means to fix (use None for those that should be estimated)
+            fixed_covariances: Covariances to fix (use None for those that should be estimated)
+            kwargs: Additional arguments passed to scikit-learn's GaussianMixture
+        """
+        super().__init__(n_components=n_components, **kwargs)
+        self.fixed_means = fixed_means
+        self.fixed_covariances = fixed_covariances
+
+        self.fixed_mean_indices = []
+        self.fixed_mean_values = []
+        if fixed_means is not None:
+            self.fixed_mean_indices = [i for i, m in enumerate(fixed_means) if m is not None]
+            if self.fixed_mean_indices:
+                self.fixed_mean_values = np.array([fixed_means[i] for i in self.fixed_mean_indices])
+
+        self.fixed_cov_indices = []
+        self.fixed_cov_values = []
+        if fixed_covariances is not None:
+            self.fixed_cov_indices = [i for i, c in enumerate(fixed_covariances) if c is not None]
+            if self.fixed_cov_indices:
+                self.fixed_cov_values = np.array([fixed_covariances[i] for i in self.fixed_cov_indices])
+
+    def _m_step(self, X: np.ndarray, log_resp: np.ndarray):
+        """Modified M-step to respect fixed means and covariances."""
+        super()._m_step(X, log_resp)
+
+        if self.fixed_mean_indices:
+            self.means_[self.fixed_mean_indices] = self.fixed_mean_values
+
+        if self.fixed_cov_indices:
+            self.covariances_[self.fixed_cov_indices] = self.fixed_cov_values
+
+        return self