spateo.external.MERFISHVI._model

Attributes

logger

Classes

SpatialVI	Single-cell Variational Inference model.
MERFISHVI	MERFISH spatial multimodal variational inference model.

Module Contents

spateo.external.MERFISHVI._model.logger = None

class spateo.external.MERFISHVI._model.SpatialVI(adata: anndata.AnnData | None = None, n_hidden: int = 128, n_latent: int = 10, n_spatial: int = 10, n_layers: int = 1, dropout_rate: float = 0.1, dispersion: Literal['gene', 'gene-batch', 'gene-label', 'gene-cell'] = 'gene', gene_likelihood: Literal['zinb', 'nb', 'poisson', 'normal'] = 'zinb', use_observed_lib_size: bool = True, latent_distribution: Literal['normal', 'ln'] = 'normal', use_spatial: bool = False, spatial_graph_type: Literal['knn', 'delaunay'] = 'knn', n_neighbors: int = 10, attention_heads: int = 1, spatial_kl_weight: float = 0.01, **kwargs)

Bases: scvi.model.base.EmbeddingMixin, scvi.model.base.RNASeqMixin, scvi.model.base.VAEMixin, scvi.model.base.ArchesMixin, scvi.model.base.UnsupervisedTrainingMixin, scvi.model.base.BaseMinifiedModeModelClass

Single-cell Variational Inference model.

This model uses a variational autoencoder (VAE) to learn low-dimensional representations of single-cell data. It can be used for batch effect correction, differential expression analysis, data imputation, and various other downstream tasks.

Parameters:

adata

AnnData object containing single-cell data, which must be registered through the setup_anndata method.

n_hidden

Number of nodes in hidden layers.

n_latent

Dimension of the latent space (dimension of low-dimensional representation).

n_spatial

Dimension of spatial features. Effective when use_spatial=True.

n_layers

Number of hidden layers in encoder and decoder networks.

dropout_rate

Ratio of neurons randomly dropped during training.

dispersion

How to model variance parameters: * ‘gene’ - one parameter per gene * ‘gene-batch’ - different parameters for each gene in each batch * ‘gene-label’ - different parameters for each gene in each label group * ‘gene-cell’ - different parameters for each gene in each cell

gene_likelihood

Distribution used to model gene expression: * ‘nb’ - negative binomial distribution (for count data with overdispersion) * ‘zinb’ - zero-inflated negative binomial distribution (for sparse count data) * ‘poisson’ - Poisson distribution (for count data) * ‘normal’ - normal distribution (experimental)

use_observed_lib_size

Whether to use observed library size as scaling factor.

latent_distribution

Type of latent distribution: * ‘normal’ - standard normal distribution * ‘ln’ - logistic normal distribution

use_spatial

Whether to use spatial information. If True, spatial coordinates will be read from adata.obsm[‘spatial’].

spatial_graph_type

Method for spatial graph construction: * ‘knn’ - K-nearest neighbors graph * ‘delaunay’ - Delaunay triangulation

n_neighbors

Number of neighbors for KNN graph when spatial_graph_type=’knn’.

attention_heads

Number of attention heads in graph attention network.

spatial_kl_weight

Weight for KL divergence of spatial latent variables.

**kwargs

Additional parameters passed to the VAE module.

Examples

>>> import anndata
>>> import scvi
>>> adata = anndata.read_h5ad("my_data.h5ad")
>>> scvi.model.SCVI.setup_anndata(adata, batch_key="batch")
>>> model = scvi.model.SCVI(adata, use_spatial=True)
>>> model.train()
>>> adata.obsm["X_scVI"] = model.get_latent_representation()

_module_cls

latent_mean_key = 'scvi_latent_qzm'

latent_var_key = 'scvi_latent_qzv'

_module_kwargs

_model_summary_string

use_spatial = False

spatial_graph_type = 'knn'

n_neighbors = 10

attention_heads = 1

spatial_kl_weight = 0.01

edge_index = None

init_params_

classmethod setup_anndata(adata: anndata.AnnData, layer: str | None = None, batch_key: str | None = None, labels_key: str | None = None, size_factor_key: str | None = None, categorical_covariate_keys: list[str] | None = None, continuous_covariate_keys: list[str] | None = None, **kwargs)

Set up AnnData object for SpatialVI model training.

This method prepares for training by registering necessary data fields.

Parameters:

adata

AnnData object containing single-cell data

layer

If provided, use this layer’s expression values instead of adata.X

batch_key

Key in adata.obs representing batch information. If None, assumes all cells are from same batch.

labels_key

Key in adata.obs representing cell type or other label information

size_factor_key

Key in adata.obs representing size factor information. Used for normalization.

categorical_covariate_keys

List of keys in adata.obs representing categorical covariates (like experimental conditions)

continuous_covariate_keys

List of keys in adata.obs representing continuous covariates (like quality control metrics)

setup_spatial_graph(adata: anndata.AnnData)

Set up spatial graph for spatial information processing.

Build spatial graph based on spatial coordinates in adata.obsm[‘spatial’], choosing between K-nearest neighbors graph or Delaunay triangulation.

Parameters:

adata

AnnData object containing spatial coordinates in adata.obsm[‘spatial’]

class spateo.external.MERFISHVI._model.MERFISHVI(adata_spatial: anndata.AnnData, adata_nonspatial: anndata.AnnData | None = None, n_hidden: int = 128, n_latent: int = 10, n_spatial: int = 10, n_layers: int = 1, dropout_rate: float = 0.1, dispersion: Literal['gene', 'gene-batch', 'gene-label', 'gene-cell'] = 'gene', gene_likelihood: Literal['zinb', 'nb', 'poisson', 'normal'] = 'zinb', use_observed_lib_size: bool = True, latent_distribution: Literal['normal', 'ln'] = 'normal', spatial_graph_type: Literal['knn', 'delaunay'] = 'knn', n_neighbors: int = 10, attention_heads: int = 1, spatial_kl_weight: float = 0.01, modality_weights: Dict[str, float] = {'spatial': 1.0, 'nonspatial': 1.0}, **kwargs)

Bases: SpatialVI

MERFISH spatial multimodal variational inference model.

This model extends SCVI to simultaneously process MERFISH data with spatial information and other modality data (such as scRNA-seq) without spatial information. It uses a shared latent space to jointly model both modalities.

Parameters:

adata_spatial

AnnData object containing spatial modality data

adata_nonspatial

AnnData object containing non-spatial modality data, optional

n_hidden

Number of nodes in hidden layers

n_latent

Dimension of latent space

n_spatial

Dimension of spatial features

n_layers

Number of hidden layers in encoder and decoder networks

dropout_rate

Dropout rate during training

dispersion

How to model variance parameters

gene_likelihood

Distribution used to model gene expression

latent_distribution

Type of latent distribution

spatial_graph_type

Spatial graph construction method, ‘knn’ or ‘delaunay’

n_neighbors

Number of neighbors when spatial_graph_type=’knn’

attention_heads

Number of attention heads in graph attention network

spatial_kl_weight

Weight for spatial feature KL divergence

modality_weights

Modality weights to balance contribution of each modality in the loss function

**kwargs

Additional parameters passed to VAE module

adata_nonspatial = None

modality_weights

_module_kwargs

_model_summary_string

spatial_graph_type = 'knn'

n_neighbors = 10

attention_heads = 1

spatial_kl_weight = 0.01

edge_index = None

init_params_

classmethod setup_nonspatial_anndata(adata: anndata.AnnData, **kwargs)

Set up non-spatial modality AnnData object.

Parameters:

adata

Non-spatial modality AnnData object

_create_module()

Create model module

get_spatial_representation(adata=None, indices=None, batch_size=None)

Get spatial feature representation

get_latent_representation(adata=None, indices=None, batch_size=None, modality='spatial')

Get latent representation, can choose from spatial modality, non-spatial modality, or fused representation

get_fused_representation(adata=None, indices=None, batch_size=None)

Get fused latent representation