spateo.tools.spatial_impute.impute#

Denoising and imputation of sparse spatial transcriptomics data

Note that this functionality requires PyTorch >= 1.8.0 Also note that this provides an alternative method for finding spatial domains (not yet fully implemented)

Module Contents#

Classes#

STGNN

Graph neural network for representation learning of spatial transcriptomics data from only the gene expression

Trainer

Graph neural network training module.

Functions#

permutation(→ torch.Tensor)

Given counts matrix in tensor form, return counts matrix with scrambled rows/spot names

get_aug_feature(adata[, highly_variable])

From AnnData object, get counts matrix, augment it and store both as .obsm entries

fix_seed([seed])

Set seeds for all random number generators using 'seed' parameter (defaults to 888)

add_contrastive_label(adata)

Creates array with 1 and 0 labels for each spot- for contrastive learning
spateo.tools.spatial_impute.impute.permutation(feature: torch.FloatTensor) → torch.Tensor[source]#

Given counts matrix in tensor form, return counts matrix with scrambled rows/spot names

spateo.tools.spatial_impute.impute.get_aug_feature(adata: anndata.AnnData, highly_variable: bool = False)[source]#

From AnnData object, get counts matrix, augment it and store both as .obsm entries

Parameters
adata

Source AnnData object

highly_variable

Set True to subset to highly-variable genes

spateo.tools.spatial_impute.impute.fix_seed(seed: int = 888)[source]#

Set seeds for all random number generators using ‘seed’ parameter (defaults to 888)

spateo.tools.spatial_impute.impute.add_contrastive_label(adata)[source]#

Creates array with 1 and 0 labels for each spot- for contrastive learning

class spateo.tools.spatial_impute.impute.STGNN(adata: anndata.AnnData, spatial_key: str = 'spatial', random_seed: int = 50, add_regularization: bool = True, device: str = 'cpu')[source]#

Graph neural network for representation learning of spatial transcriptomics data from only the gene expression matrix. Wraps preprocessing and training.

adata: class anndata.AnnData spatial_key: Key in .obsm where x- and y-coordinates are stored random_seed: Sets seed for all random number generators add_regularization: Set True to include weight-based penalty term in representation learning. device: Options: ‘cpu’, ‘cuda:_’. Perform computations on CPU or GPU. If GPU, provide the name of the device to run

computations

train_STGNN(**kwargs)[source]#
Parameters
kwargs

Arguments that can be passed to :class Trainer.

Returns

AnnData object with the smoothed values stored in a layer, either “X_smooth_gcn” or

”X_smooth_gcn_reg”.

Return type

adata_output

class spateo.tools.spatial_impute.impute.Trainer(adata: anndata.AnnData, device: torch.device, learn_rate: float = 0.001, dropout: float = 0.0, act=F.relu, clip: Union[None, float] = 0.25, weight_decay: float = 0.0, epochs: int = 1000, dim_output: int = 64, alpha: float = 10, beta: float = 1, theta: float = 0.1, add_regularization: bool = False)[source]#

Graph neural network training module.

Parameters
adata

class anndata.AnnData

device

torch.device object

learn_rate

Controls magnitude of gradient for network learning

dropout

Proportion of weights in each layer to set to 0

act

String specifying activation function for each encoder layer. Options: “linear”, “sigmoid”, “tanh”, “relu”, “elu”

clip

Threshold below which imputed feature values will be set to 0, as a percentile

weight_decay

Controls degradation rate of parameters

epochs

Number of iterations of training loop to perform

dim_output

Dimensionality of the output representation

alpha

Controls influence of reconstruction loss in representation learning

beta

Weight factor to control the influence of contrastive loss in representation learning

theta

Weight factor to control the influence of the regularization term in representation learning

add_regularization

Adds penalty term to representation learning

train()[source]#
Returns

-emb_rec (ndarray) Reconstruction of the counts matrix