scanpy.external.pp.dca(adata, mode='denoise', ae_type='nb-conddisp', normalize_per_cell=True, scale=True, log1p=True, hidden_size=(64, 32, 64), hidden_dropout=0.0, batchnorm=True, activation='relu', init='glorot_uniform', network_kwds=mappingproxy({}), epochs=300, reduce_lr=10, early_stop=15, batch_size=32, optimizer='RMSprop', random_state=0, threads=None, learning_rate=None, verbose=False, training_kwds=mappingproxy({}), return_model=False, return_info=False, copy=False)

Deep count autoencoder [Eraslan18].

Fits a count autoencoder to the raw count data given in the anndata object in order to denoise the data and to capture hidden representation of cells in low dimensions. Type of the autoencoder and return values are determined by the parameters.


More information and bug reports here.

adata : AnnData

An anndata file with .raw attribute representing raw counts.

mode : Literal[‘denoise’, ‘latent’] (default: 'denoise')

denoise overwrites adata.X with denoised expression values. In latent mode DCA adds adata.obsm['X_dca'] to given adata object. This matrix represent latent representation of cells via DCA.

ae_type : Literal[‘zinb-conddisp’, ‘zinb’, ‘nb-conddisp’, ‘nb’] (default: 'nb-conddisp')

Type of the autoencoder. Return values and the architecture is determined by the type e.g. nb does not provide dropout probabilities. Types that end with “-conddisp”, assumes that dispersion is mean dependant.

normalize_per_cell : bool (default: True)

If true, library size normalization is performed using the sc.pp.normalize_per_cell function in Scanpy and saved into adata object. Mean layer is re-introduces library size differences by scaling the mean value of each cell in the output layer. See the manuscript for more details.

scale : bool (default: True)

If true, the input of the autoencoder is centered using sc.pp.scale function of Scanpy. Note that the output is kept as raw counts as loss functions are designed for the count data.

log1p : bool (default: True)

If true, the input of the autoencoder is log transformed with a pseudocount of one using sc.pp.log1p function of Scanpy.

hidden_size : Sequence[int] (default: (64, 32, 64))

Width of hidden layers.

hidden_dropout : Union[float, Sequence[float]] (default: 0.0)

Probability of weight dropout in the autoencoder (per layer if list or tuple).

batchnorm : bool (default: True)

If true, batch normalization is performed.

activation : str (default: 'relu')

Activation function of hidden layers.

init : str (default: 'glorot_uniform')

Initialization method used to initialize weights.

network_kwds : Mapping[str, Any] (default: mappingproxy({}))

Additional keyword arguments for the autoencoder.

epochs : int (default: 300)

Number of total epochs in training.

reduce_lr : int (default: 10)

Reduces learning rate if validation loss does not improve in given number of epochs.

early_stop : int (default: 15)

Stops training if validation loss does not improve in given number of epochs.

batch_size : int (default: 32)

Number of samples in the batch used for SGD.

optimizer : str (default: 'RMSprop')

Type of optimization method used for training.

random_state : Union[None, int, RandomState] (default: 0)

Seed for python, numpy and tensorflow.

threads : Optional[int] (default: None)

Number of threads to use in training. All cores are used by default.

learning_rate : Optional[float] (default: None)

Learning rate to use in the training.

verbose : bool (default: False)

If true, prints additional information about training and architecture.

training_kwds : Mapping[str, Any] (default: mappingproxy({}))

Additional keyword arguments for the training process.

return_model : bool (default: False)

If true, trained autoencoder object is returned. See “Returns”.

return_info : bool (default: False)

If true, all additional parameters of DCA are stored in adata.obsm such as dropout probabilities (obsm[‘X_dca_dropout’]) and estimated dispersion values (obsm[‘X_dca_dispersion’]), in case that autoencoder is of type zinb or zinb-conddisp.

copy : bool (default: False)

If true, a copy of anndata is returned.

Return type



If copy is true and return_model is false, AnnData object is returned.

In “denoise” mode, adata.X is overwritten with the denoised values. In “latent” mode, latent low dimensional representation of cells are stored in adata.obsm['X_dca'] and adata.X is not modified. Note that these values are not corrected for library size effects.

If return_info is true, all estimated distribution parameters are stored in AnnData like this:


The mixture coefficient (pi) of the zero component in ZINB, i.e. dropout probability (if ae_type is zinb or zinb-conddisp).


The dispersion parameter of NB.


The loss history of the training. See .history attribute of Keras History class for mode details.

Finally, the raw counts are stored in .raw attribute of AnnData object.

If return_model is given, trained model is returned. When both copy and return_model are true, a tuple of anndata and model is returned in that order.