pertpy.tools.MLPClassifierSpace

pertpy.tools.MLPClassifierSpace#

class MLPClassifierSpace[source]#

Fits an ANN classifier to the data and takes the feature space (weights in the last layer) as embedding.

We train the ANN to classify the different perturbations. After training, the penultimate layer is used as the feature space, resulting in one embedding per cell. Consider employing the PseudoBulk or another PerturbationSpace to obtain one embedding per perturbation.

See here https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7289078/ (Dose-response analysis) and Sup 17-19.

Methods table#

add(adata, *, perturbations[, ...])

Add perturbations linearly.

compute(adata[, target_col, layer_key, ...])

Creates a perturbation embedding by training a MLP classifier model to distinguish between perturbations.

compute_control_diff(adata, *[, target_col, ...])

Subtract mean of the control from the perturbation.

dose_response(adata, *[, target_col, ...])

Quantify the effect size of each perturbation as a function of dose.

evaluate_combinations(adata, *[, ...])

Score how well an additive model predicts combination perturbations.

label_transfer(adata, *[, target_column, ...])

Impute missing values in the specified column using KNN imputation in the space defined by use_rep.

nearest_perturbations(adata, perturbation, *)

Rank perturbations by their proximity to a query perturbation in a perturbation space.

plot_similarity(adata, *[, target_col, ...])

Plot a clustered heatmap of pairwise distances between perturbations.

subtract(adata, *, perturbations[, ...])

Subtract perturbations linearly.

Methods#

MLPClassifierSpace.add(adata, *, perturbations, reference_key='control', ensure_consistency=True, target_col='perturbation')#

Add perturbations linearly. Assumes input of size n_perts x dimensionality.

Parameters:

  • adata (AnnData) – Anndata object of size n_perts x dim.

  • perturbations (Iterable[str]) – Perturbations to add.

  • reference_key (str, default: 'control') – perturbation source from which the perturbation summation starts.

  • ensure_consistency (bool, default: True) – If True, differentiate against control via compute_control_diff before combining so that “perturbation - perturbation == control” holds in the resulting space. Set False only if the input has already been differenced.

  • target_col (str, default: 'perturbation') – .obs column name that stores the label of the perturbation applied to each cell.

Return type:

tuple[AnnData, AnnData] | AnnData

Returns:

Anndata object of size (n_perts+1) x dim, where the last row is the addition of the specified perturbations. If ensure_consistency is True, returns a tuple of (new_perturbation, adata) where adata is the AnnData object provided as input but updated using compute_control_diff.

Examples

Example usage with PseudobulkSpace:

>>> import pertpy as pt
>>> mdata = pt.dt.papalexi_2021()
>>> ps = pt.tl.PseudobulkSpace()
>>> ps_adata = ps.compute(mdata["rna"], target_col="gene_target", groups_col="gene_target")
>>> new_perturbation = ps.add(ps_adata, perturbations=["ATF2", "CD86"], reference_key="NT")
MLPClassifierSpace.compute(adata, target_col='perturbation', layer_key=None, embedding_key=None, hidden_dim=None, dropout=0.0, batch_norm=True, batch_size=128, test_split_size=0.2, validation_split_size=0.25, max_epochs=20, val_epochs_check=2, patience=2, lr=0.0001, seed=42)[source]#

Creates a perturbation embedding by training a MLP classifier model to distinguish between perturbations.

A model is created using the specified parameters (hidden_dim, dropout, batch_norm). Further parameters such as the number of classes to predict (number of perturbations) are obtained from the provided AnnData object directly. Dataloaders that take into account class imbalances are created. Next, the model is trained and tested, using the GPU if available. The penultimate-layer activations are extracted for every cell and averaged per perturbation, yielding one embedding per perturbation.

Parameters:

  • adata (AnnData) – AnnData object of size cells x genes

  • target_col (str, default: 'perturbation') – .obs column that stores the label of the perturbation applied to each cell.

  • layer_key (str | None, default: None) – Layer in adata to use.

  • embedding_key (str | None, default: None) – .obsm embedding to train the classifier on. Mutually exclusive with layer_key.

  • hidden_dim (list[int] | None, default: None) – List of number of neurons in each hidden layers of the neural network. For instance, [512, 256] will create a neural network with two hidden layers, the first with 512 neurons and the second with 256 neurons.

  • dropout (float, default: 0.0) – Amount of dropout applied, constant for all layers.

  • batch_norm (bool, default: True) – Whether to apply batch normalization.

  • batch_size (int, default: 128) – The batch size, i.e. the number of datapoints to use in one forward/backward pass.

  • test_split_size (float, default: 0.2) – Fraction of data to put in the test set. Default to 0.2.

  • validation_split_size (float, default: 0.25) – Fraction of data to put in the validation set of the resultant train set. E.g. a test_split_size of 0.2 and a validation_split_size of 0.25 means that 25% of 80% of the data will be used for validation.

  • max_epochs (int, default: 20) – Maximum number of epochs for training.

  • val_epochs_check (int, default: 2) – Test performance on validation dataset after every val_epochs_check training epochs. Note that this affects early stopping, as the model will be stopped if the validation performance does not improve for patience epochs.

  • patience (int, default: 2) – Number of validation performance checks without improvement, after which the early stopping flag is activated and training is therefore stopped.

  • lr (float, default: 0.0001) – Learning rate for training.

  • seed (int, default: 42) – Random seed for reproducibility.

Return type:

AnnData

Returns:

AnnData with one observation per perturbation, the averaged penultimate-layer embedding in .X and the perturbation labels in .obs[target_col].

Examples

>>> import pertpy as pt
>>> adata = pt.dt.norman_2019()
>>> dcs = pt.tl.MLPClassifierSpace()
>>> pert_embeddings = dcs.compute(adata, target_col="perturbation_name")
MLPClassifierSpace.compute_control_diff(adata, *, target_col='perturbation', group_col=None, reference_key='control', layer_key=None, new_layer_key='control_diff', embedding_key=None, new_embedding_key='control_diff', all_data=False, copy=True)#

Subtract mean of the control from the perturbation.

Parameters:

  • adata (AnnData) – Anndata object of size cells x genes.

  • target_col (str, default: 'perturbation') – .obs column name that stores the label of the perturbation applied to each cell.

  • group_col (str | None, default: None) – .obs column name that stores the label of the group of each cell. If None, ignore groups.

  • reference_key (str, default: 'control') – The key of the control values.

  • layer_key (str | None, default: None) – Key of the AnnData layer to use for computation.

  • new_layer_key (str, default: 'control_diff') – the results are stored in the given layer.

  • embedding_key (str | None, default: None) – obsm key of the AnnData embedding to use for computation.

  • new_embedding_key (str, default: 'control_diff') – Results are stored in a new embedding in obsm with this key.

  • all_data (bool, default: False) – if True, do the computation in all data representations (X, all layers and all embeddings)

  • copy (bool, default: True) – If True returns a new AnnData; otherwise updates the input AnnData in place.

Return type:

AnnData

Returns:

Updated AnnData object.

Examples

Example usage with PseudobulkSpace:

>>> import pertpy as pt
>>> mdata = pt.dt.papalexi_2021()
>>> ps = pt.tl.PseudobulkSpace()
>>> diff_adata = ps.compute_control_diff(mdata["rna"], target_col="gene_target", reference_key="NT")
MLPClassifierSpace.dose_response(adata, *, target_col='perturbation', dose_col='dose', reference_key='control', metric='edistance', layer_key=None, embedding_key=None, **kwargs)#

Quantify the effect size of each perturbation as a function of dose.

For every (perturbation, dose) group the statistical distance to reference_key is computed in the chosen representation using Distance. Operates on cell-level data, since distances are defined between groups of cells.

Parameters:

  • adata (AnnData) – Cell-level AnnData.

  • target_col (str, default: 'perturbation') – .obs column with the perturbation label.

  • dose_col (str, default: 'dose') – .obs column with the (numeric) dose.

  • reference_key (str, default: 'control') – Control perturbation all doses are compared against.

  • metric (Literal['edistance', 'euclidean', 'root_mean_squared_error', 'mse', 'mean_absolute_error', 'pearson_distance', 'spearman_distance', 'kendalltau_distance', 'cosine_distance', 'r2_distance', 'mean_pairwise', 'mmd', 'wasserstein', 'sym_kldiv', 't_test', 'ks_test', 'nb_ll', 'classifier_proba', 'classifier_cp', 'mean_var_distribution', 'mahalanobis'], default: 'edistance') – Distance metric passed to Distance.

  • layer_key (str | None, default: None) – Layer to compute distances from.

  • embedding_key (str | None, default: None) – .obsm embedding to compute distances from.

  • kwargs – Passed to onesided_distances().

Return type:

DataFrame

Returns:

Tidy DataFrame with perturbation, dose and distance columns, sorted by perturbation then dose.

Examples

>>> import pertpy as pt
>>> adata = pt.dt.srivatsan_2020_sciplex2()
>>> ps = pt.tl.PseudobulkSpace()
>>> curves = ps.dose_response(adata, dose_col="dose_value", embedding_key="X_pca")
MLPClassifierSpace.evaluate_combinations(adata, *, combinations=None, target_col='perturbation', reference_key='control', metric='pearson', sep='+')#

Score how well an additive model predicts combination perturbations.

For every combination "A+B" whose components A and B are each present as single perturbations, the additive prediction effect(A) + effect(B) is compared against the measured combination effect, where effects are taken relative to reference_key. A small distance indicates additive, non-interacting perturbations, whereas a large deviation flags a genetic or pharmacological interaction.

Parameters:

  • adata (AnnData) – Perturbation-level AnnData indexed by perturbation (one observation per perturbation).

  • combinations (Iterable[str] | None, default: None) – Combination names to evaluate. If None, every obs_name containing sep whose components are all present as singles is used.

  • target_col (str, default: 'perturbation') – .obs column identifying each perturbation.

  • reference_key (str, default: 'control') – Control perturbation subtracted to obtain effects. If absent, values are used as-is.

  • metric (Literal['pearson', 'cosine', 'euclidean'], default: 'pearson') – Distance between predicted and measured combination effects.

  • sep (str, default: '+') – Separator between components in combination names.

Return type:

DataFrame

Returns:

DataFrame indexed by combination with distance, predicted_magnitude and measured_magnitude columns, sorted by ascending distance.

Examples

>>> import pertpy as pt
>>> adata = pt.dt.norman_2019()
>>> ps = pt.tl.PseudobulkSpace()
>>> ps_adata = ps.compute(adata, target_col="perturbation_name")
>>> scores = ps.evaluate_combinations(ps_adata, target_col="perturbation_name", reference_key="control")
MLPClassifierSpace.label_transfer(adata, *, target_column='perturbation', column_uncertainty_score_key='perturbation_transfer_uncertainty', target_val='unknown', neighbors_key='neighbors')#

Impute missing values in the specified column using KNN imputation in the space defined by use_rep.

Uncertainty is calculated as the entropy of the label distribution in the neighborhood of the target cell. In other words, a cell where all neighbors have the same set of labels will have an uncertainty of 0, whereas a cell where all neighbors have many different labels will have high uncertainty.

Parameters:

  • adata (AnnData) – The AnnData object containing single-cell data.

  • target_column (str, default: 'perturbation') – The column name in adata.obs to perform imputation on.

  • column_uncertainty_score_key (str, default: 'perturbation_transfer_uncertainty') – The column name in adata.obs to store the uncertainty score of the label transfer.

  • target_val (str, default: 'unknown') – The target value to impute.

  • neighbors_key (str, default: 'neighbors') – The key in adata.uns where the neighbors are stored.

Return type:

None

Examples

>>> import pertpy as pt
>>> import scanpy as sc
>>> import numpy as np
>>> adata = sc.datasets.pbmc68k_reduced()
>>> # randomly dropout 10% of the data annotations
>>> adata.obs["perturbation"] = adata.obs["louvain"].astype(str).copy()
>>> random_cells = np.random.choice(adata.obs.index, int(adata.obs.shape[0] * 0.1), replace=False)
>>> adata.obs.loc[random_cells, "perturbation"] = "unknown"
>>> sc.pp.neighbors(adata)
>>> sc.tl.umap(adata)
>>> ps = pt.tl.PseudobulkSpace()
>>> ps.label_transfer(adata)
MLPClassifierSpace.nearest_perturbations(adata, perturbation, *, target_col='perturbation', n_neighbors=10, layer_key=None, embedding_key=None, metric='euclidean')#

Rank perturbations by their proximity to a query perturbation in a perturbation space.

Operates on a perturbation-level AnnData (one observation per perturbation), i.e. the output of any compute. Useful for discovering perturbations with a similar mechanism of action. If adata.obsp["distances"] is present (as produced by DistanceSpace) and no representation is requested explicitly, those precomputed distances are used directly.

Parameters:

  • adata (AnnData) – Perturbation-level AnnData indexed by perturbation.

  • perturbation (str) – The query perturbation to find neighbors for.

  • target_col (str, default: 'perturbation') – .obs column identifying each perturbation.

  • n_neighbors (int, default: 10) – Number of nearest perturbations to return.

  • layer_key (str | None, default: None) – Layer to compute distances from.

  • embedding_key (str | None, default: None) – .obsm embedding to compute distances from.

  • metric (str, default: 'euclidean') – Distance metric passed to sklearn.metrics.pairwise_distances().

Return type:

DataFrame

Returns:

DataFrame indexed by perturbation with a distance column, sorted ascending and excluding the query.

Examples

>>> import pertpy as pt
>>> adata = pt.dt.norman_2019()
>>> ps = pt.tl.PseudobulkSpace()
>>> ps_adata = ps.compute(adata, target_col="perturbation_name")
>>> neighbors = ps.nearest_perturbations(ps_adata, "CBL+CNN1", target_col="perturbation_name")
MLPClassifierSpace.plot_similarity(adata, *, target_col='perturbation', layer_key=None, embedding_key=None, metric='euclidean', cmap='viridis', **kwargs)#

Plot a clustered heatmap of pairwise distances between perturbations.

Uses adata.obsp["distances"] when present (e.g. from DistanceSpace) and otherwise computes pairwise distances in the chosen representation.

Parameters:

  • adata (AnnData) – Perturbation-level AnnData indexed by perturbation.

  • target_col (str, default: 'perturbation') – .obs column identifying each perturbation.

  • layer_key (str | None, default: None) – Layer to compute distances from.

  • embedding_key (str | None, default: None) – .obsm embedding to compute distances from.

  • metric (str, default: 'euclidean') – Distance metric passed to sklearn.metrics.pairwise_distances().

  • cmap (str, default: 'viridis') – Matplotlib colormap.

  • kwargs – Passed to seaborn.clustermap().

Returns:

The grid returned by seaborn.clustermap().

Examples

>>> import pertpy as pt
>>> adata = pt.dt.norman_2019()
>>> ds = pt.tl.DistanceSpace()
>>> ds_adata = ds.compute(adata, target_col="perturbation_name", metric="edistance")
>>> ds.plot_similarity(ds_adata, target_col="perturbation_name")
MLPClassifierSpace.subtract(adata, *, perturbations, reference_key='control', ensure_consistency=True, target_col='perturbation')#

Subtract perturbations linearly. Assumes input of size n_perts x dimensionality.

Parameters:

  • adata (AnnData) – Anndata object of size n_perts x dim.

  • perturbations (Iterable[str]) – Perturbations to subtract.

  • reference_key (str, default: 'control') – Perturbation source from which the perturbation subtraction starts.

  • ensure_consistency (bool, default: True) – If True, differentiate against control via compute_control_diff before combining so that “perturbation - perturbation == control” holds in the resulting space. Set False only if the input has already been differenced.

  • target_col (str, default: 'perturbation') – .obs column name that stores the label of the perturbation applied to each cell.

Return type:

tuple[AnnData, AnnData] | AnnData

Returns:

Anndata object of size (n_perts+1) x dim, where the last row is the subtraction of the specified perturbations. If ensure_consistency is True, returns a tuple of (new_perturbation, adata) where adata is the AnnData object provided as input but updated using compute_control_diff.

Examples

Example usage with PseudobulkSpace:

>>> import pertpy as pt
>>> mdata = pt.dt.papalexi_2021()
>>> ps = pt.tl.PseudobulkSpace()
>>> ps_adata = ps.compute(mdata["rna"], target_col="gene_target", groups_col="gene_target")
>>> new_perturbation = ps.subtract(ps_adata, reference_key="ATF2", perturbations=["BRD4", "CUL3"])