scvi.external.DRVI

Contents

scvi.external.DRVI#

class scvi.external.DRVI(adata=None, registry=None, n_latent=32, n_split_latent=None, split_method='split_map', split_aggregation='logsumexp', gene_likelihood='pnb', **kwargs)[source]#

Disentangled Representation Variational Inference [Moinfar and Theis, 2024].

DRVI is an unsupervised deep generative model that learns an interpretable, disentangled latent representation of single-cell omics data. Disentanglement is induced in the decoder: the latent is split into independent groups, each decoded separately and aggregated (see DecoderDRVI).

DRVI is built directly on SCVI: it reuses SCVI’s encoder, likelihoods, covariate/batch handling, minified-mode, scArches and out-of-core (datamodule) machinery unchanged, and only swaps the decoder for the additive split decoder and adds the disentanglement/interpretability logic. It therefore exposes the full SCVI interface (setup_anndata(), minify_adata, get_batch_representation, size_factor_key, …) plus the DRVI-specific knobs below and the interpretability methods from InterpretabilityMixin.

Parameters:

  • adata (AnnData | None (default: None)) – AnnData object registered via setup_anndata(). May be None when initializing from a registry (e.g. out-of-core training with a datamodule) or to defer module initialization to train (datamodule path).

  • registry (dict | None (default: None)) – Setup registry (e.g. from a datamodule such as AnnbatchDataModule) to initialize without an in-memory AnnData.

  • n_latent (int (default: 32)) – Dimensionality of the latent space.

  • n_split_latent (int | None (default: None)) – Number of latent splits. None (default) splits every latent dimension.

  • split_method (Literal['split_mask', 'split_map'] (default: 'split_map')) – Latent-to-split mapping, "split_mask" or "split_map" (default).

  • split_aggregation (Literal['mean', 'logsumexp'] (default: 'logsumexp')) – Per-split aggregation, "mean" or "logsumexp" (default).

  • gene_likelihood (Literal['nb', 'pnb', 'zinb', 'poisson', 'normal', 'normal_unit_var'] (default: 'pnb')) – Reconstruction likelihood; "pnb" (default, log-space parametrized NB), "nb", "zinb", "poisson", "normal" or "normal_unit_var".

  • **kwargs – Additional keyword args for SCVI / DRVIModule (e.g. n_hidden, n_layers, n_split_output, dispersion, batch_representation, activation_fn, use_observed_lib_size).

Examples

>>> adata = anndata.read_h5ad(path_to_anndata)
>>> scvi.external.DRVI.setup_anndata(adata, batch_key="batch")
>>> model = scvi.external.DRVI(adata, n_latent=32)
>>> model.train()
>>> adata.obsm["X_drvi"] = model.get_latent_representation()

Attributes table#

adata

Data attached to model instance.

adata_manager

Manager instance associated with self.adata.

device

The current device that the module's params are on.

get_normalized_function_name

What the get normalized functions name is

history

Returns computed metrics during training.

is_trained

Whether the model has been trained.

minified_data_type

The type of minified data associated with this model, if applicable.

registry

Data attached to model instance.

run_id

Returns the run id of the model.

run_name

Returns the run name of the model.

summary_string

Summary string of the model.

test_indices

Observations that are in test set.

train_indices

Observations that are in train set.

validation_indices

Observations that are in validation set.

Methods table#

calculate_interpretability_scores(embed[, ...])

Compute per-factor, per-gene interpretability scores and (optionally) store them.

convert_legacy_save(dir_path, output_dir_path)

Converts a legacy saved model (<v0.15.0) to the updated save format.

data_registry(registry_key)

Returns the object in AnnData associated with the key in the data registry.

deregister_manager([adata])

Deregisters the AnnDataManager instance associated with adata.

differential_abundance([adata, adata_sub, ...])

Compute the differential abundance between samples.

differential_expression([adata, groupby, ...])

A unified method for differential expression analysis.

get_aggregated_posterior([adata, indices, ...])

Compute the aggregated posterior over the u latent representations.

get_anndata_manager(adata[, required])

Retrieves the AnnDataManager for a given AnnData object.

get_batch_representation([adata, indices, ...])

Get the batch representation for a given set of indices.

get_effect_of_splits_out_of_distribution(embed)

Out-of-distribution per-split, per-gene effect scores via latent traversal.

get_effect_of_splits_within_distribution([...])

In-distribution per-split, per-gene effect scores.

get_elbo([adata, indices, batch_size, ...])

Compute the evidence lower bound (ELBO) on the data.

get_feature_correlation_matrix([adata, ...])

Generate gene-gene correlation matrix using scvi uncertainty and expression.

get_from_registry(adata, registry_key)

Returns the object in AnnData associated with the key in the data registry.

get_importance_weights(adata, indices, qz, ...)

Computes importance weights for the given samples.

get_interpretability_scores(embed, adata[, ...])

Return interpretability scores as a genes (rows) × dimensions (cols) DataFrame.

get_latent_library_size([adata, indices, ...])

Returns the latent library size for each cell.

get_latent_representation([adata, indices, ...])

Compute the latent representation of the data.

get_likelihood_parameters([adata, indices, ...])

Estimates for the parameters of the likelihood \(p(x \mid z)\).

get_marginal_ll([adata, indices, ...])

Compute the marginal log-likehood of the data.

get_normalized_expression([adata, indices, ...])

Returns the normalized (decoded) gene expression.

get_reconstruction_effect_of_each_split([...])

Effect of each split on reconstruction (summed over genes).

get_reconstruction_error([adata, indices, ...])

Compute the reconstruction error on the data.

get_setup_arg(setup_arg)

Returns the string provided to setup of a specific setup_arg.

get_state_registry(registry_key)

Returns the state registry for the AnnDataField registered with this instance.

get_var_names([legacy_mudata_format])

Variable names of input data.

iterate_on_ae_output([adata, dataloader, ...])

Yield (inference_outputs, generative_outputs) per minibatch.

iterate_on_decoded_latent_samples(z[, lib, ...])

Yield decoder (generative) outputs for given latent samples, batched.

iterate_on_effect_of_splits_within_distribution([...])

Yield (effect_tensor, latent) per minibatch: each split's effect on every gene.

load(dir_path[, adata, accelerator, device, ...])

Instantiate a model from the saved output.

load_query_data([adata, reference_model, ...])

Online update of a reference model with scArches algorithm [Lotfollahi et al., 2021].

load_registry(dir_path[, prefix])

Return the full registry saved with the model.

minify_adata([minified_data_type, ...])

Minify the model's adata.

posterior_predictive_sample([adata, ...])

Generate predictive samples from the posterior predictive distribution.

prepare_query_anndata(adata, reference_model)

Prepare data for query integration.

prepare_query_mudata(mdata, reference_model)

Prepare multimodal dataset for query integration.

register_manager(adata_manager)

Registers an AnnDataManager instance with this model class.

save(dir_path[, prefix, overwrite, ...])

Save the state of the model.

set_latent_dimension_stats(embed[, adata, ...])

Annotate embed.var with per-dimension reconstruction effect, ordering and stats.

setup_annbatch(cls[, collection_path, ...])

Set up a model from on-disk AnnData files via the annbatch streaming loader.

setup_anndata(adata[, layer, batch_key, ...])

Sets up the AnnData object for this model.

to_device(device)

Move the model to the device.

train([max_epochs, accelerator, devices, ...])

Train the model.

transfer_fields(adata, **kwargs)

Transfer fields from a model to an AnnData object.

update_setup_method_args(setup_method_args)

Update setup method args.

view_anndata_setup([adata, ...])

Print summary of the setup for the initial AnnData or a given AnnData object.

view_registry([hide_state_registries])

Prints summary of the registry.

view_setup_args(dir_path[, prefix])

Print args used to setup a saved model.

view_setup_method_args()

Prints setup kwargs used to produce a given registry.

Attributes#

DRVI.adata[source]#

Data attached to model instance.

DRVI.adata_manager[source]#

Manager instance associated with self.adata.

DRVI.device[source]#

The current device that the module’s params are on.

DRVI.get_normalized_function_name[source]#

What the get normalized functions name is

DRVI.history[source]#

Returns computed metrics during training.

DRVI.is_trained[source]#

Whether the model has been trained.

DRVI.minified_data_type[source]#

The type of minified data associated with this model, if applicable.

DRVI.registry[source]#

Data attached to model instance.

DRVI.run_id[source]#

Returns the run id of the model. Used in MLFlow

DRVI.run_name[source]#

Returns the run name of the model. Used in MLFlow

DRVI.summary_string[source]#

Summary string of the model.

DRVI.test_indices[source]#

Observations that are in test set.

DRVI.train_indices[source]#

Observations that are in train set.

DRVI.validation_indices[source]#

Observations that are in validation set.

Methods#

DRVI.calculate_interpretability_scores(embed, methods='OOD', directional=True, add_to_counts=1.0, inplace=True, **kwargs)[source]#

Compute per-factor, per-gene interpretability scores and (optionally) store them.

Top-level entry point for DRVI’s latent-factor interpretability. It runs the in- and/or out-of-distribution analyses and collects their per-aggregation score matrices (one value per latent factor and gene) under a single, consistently named set of keys.

Parameters:

  • embed (AnnData) – AnnData of the latent space (one var per latent factor), as annotated by set_latent_dimension_stats(). Scores are written to its varm when inplace.

  • methods (Sequence[str] | str (default: 'OOD')) – Which analyses to run: "IND" (in-distribution, over the data via get_effect_of_splits_within_distribution()), "OOD" (default; out-of- distribution latent traversal via get_effect_of_splits_out_of_distribution()), or "ALL" for both. A sequence of these is also accepted.

  • directional (bool (default: True)) – If True (default), each factor’s positive and negative directions are scored separately and the resulting keys gain a _positive / _negative suffix. Requires n_split_latent == n_latent.

  • add_to_counts (float (default: 1.0)) – Pseudo-count forwarded to the underlying effect computations (see those methods).

  • inplace (bool (default: True)) – If True (default), store each score matrix in embed.varm and return None; otherwise return them as a dict and leave embed untouched.

  • kwargs (Any) – Extra keyword arguments forwarded to the underlying method(s), each filtered to the arguments that method actually accepts. Useful ones include adata / dataloader, deterministic and skip_threshold (IND), and n_steps, n_samples, seed and batch_size (OOD).

Return type:

dict[str, ndarray] | None

Returns:

None when inplace=True (scores written to embed.varm), otherwise a dict mapping each key to its (n_factors, n_genes) score matrix.

Notes

Keys are "{method}_{aggregation}[_positive|_negative]". The aggregations are max, linear_weighted_mean and exp_weighted_mean for IND, and min_possible, max_possible and combined for OOD (e.g. "OOD_combined_positive"). These are the keys get_interpretability_scores() reads back (default "OOD_combined").

classmethod DRVI.convert_legacy_save(dir_path, output_dir_path, overwrite=False, prefix=None, **save_kwargs)[source]#

Converts a legacy saved model (<v0.15.0) to the updated save format.

Parameters:

  • dir_path (str) – Path to the directory where the legacy model is saved.

  • output_dir_path (str) – Path to save converted save files.

  • overwrite (bool (default: False)) – Overwrite existing data or not. If False and directory already exists at output_dir_path, an error will be raised.

  • prefix (str | None (default: None)) – Prefix of saved file names.

  • **save_kwargs – Keyword arguments passed into save().

Return type:

None

DRVI.data_registry(registry_key)[source]#

Returns the object in AnnData associated with the key in the data registry.

Parameters:

registry_key (str) – key of an object to get from self.data_registry

Return type:

ndarray | DataFrame

Returns:

The requested data.

DRVI.deregister_manager(adata=None)[source]#

Deregisters the AnnDataManager instance associated with adata.

If adata is None, deregisters all AnnDataManager instances in both the class and instance-specific manager stores, except for the one associated with this model instance.

DRVI.differential_abundance(adata=None, adata_sub=None, sample_key=None, batch_size=128, num_cells_posterior=None, dof=None, dataloader=None)[source]#

Compute the differential abundance between samples.

Computes the log probabilities of each sample conditioned on the estimated aggregate posterior distribution of each cell.

Parameters:

  • adata (AnnData | MuData | None (default: None)) – The full data object used to compute each aggregated posterior. Defaults to the AnnData object used to initialize the model.

  • adata_sub (AnnData | MuData | None (default: None)) – The data object to compute the differential abundance for. For very large datasets, this should be used to pass in a subset of the full data object. The aggregated posteriors are still computed from the full data object. The resulting log_probs matrix is stored in adata_sub.obsm

  • sample_key (str | None (default: None)) – Key for the sample covariate.

  • batch_size (int (default: 128)) – Minibatch size for computing the differential abundance.

  • num_cells_posterior (int | None (default: None)) – Maximum number of cells used to compute aggregated posterior for each sample.

  • dof (float | None (default: None)) – Degrees of freedom for the Student’s t-distribution components for aggregated posterior. If None, components are Normal.

  • dataloader (Iterator[dict[str, Tensor | None]] | None (default: None)) – Inference dataloader to materialize when the model was initialized without AnnData.

DRVI.differential_expression(adata=None, groupby=None, group1=None, group2=None, idx1=None, idx2=None, mode='vanilla', delta=0.25, batch_size=None, all_stats=True, batch_correction=False, batchid1=None, batchid2=None, fdr_target=0.05, silent=False, weights='uniform', filter_outlier_cells=False, importance_weighting_kwargs=None, dataloader=None, **kwargs)[source]#

A unified method for differential expression analysis.

Implements 'vanilla' DE [Lopez et al., 2018] and 'change' mode DE [Boyeau et al., 2019].

Parameters:

  • adata (AnnData | None (default: None)) – AnnData object with equivalent structure to initial AnnData. If None, defaults to the AnnData object used to initialize the model.

  • groupby (str | None (default: None)) – The key of the observations grouping to consider.

  • group1 (list[str] | None (default: None)) – Subset of groups, e.g. [‘g1’, ‘g2’, ‘g3’], to which comparison shall be restricted, or all groups in groupby (default).

  • group2 (str | None (default: None)) – If None, compare each group in group1 to the union of the rest of the groups in groupby. If a group identifier, compare with respect to this group.

  • idx1 (list[int] | list[bool] | str | None (default: None)) – idx1 and idx2 can be used as an alternative to the AnnData keys. Custom identifier for group1 that can be of three sorts: (1) a boolean mask, (2) indices, or (3) a string. If it is a string, then it will query indices that verifies conditions on adata.obs, as described in pandas.DataFrame.query() If idx1 is not None, this option overrides group1 and group2.

  • idx2 (list[int] | list[bool] | str | None (default: None)) – Custom identifier for group2 that has the same properties as idx1. By default, includes all cells not specified in idx1.

  • mode (Literal['vanilla', 'change'] (default: 'vanilla')) – Method for differential expression. See user guide for full explanation.

  • delta (float (default: 0.25)) – specific case of region inducing differential expression. In this case, we suppose that \(R \setminus [-\delta, \delta]\) does not induce differential expression (change model default case).

  • batch_size (int | None (default: None)) – Minibatch size for data loading into model. Defaults to scvi.settings.batch_size.

  • all_stats (bool (default: True)) – Concatenate count statistics (e.g., mean expression group 1) to DE results.

  • batch_correction (bool (default: False)) – Whether to correct for batch effects in DE inference.

  • batchid1 (list[str] | None (default: None)) – Subset of categories from batch_key registered in setup_anndata, e.g. [‘batch1’, ‘batch2’, ‘batch3’], for group1. Only used if batch_correction is True, and by default all categories are used.

  • batchid2 (list[str] | None (default: None)) – Same as batchid1 for group2. batchid2 must either have null intersection with batchid1, or be exactly equal to batchid1. When the two sets are exactly equal, cells are compared by decoding on the same batch. When sets have null intersection, cells from group1 and group2 are decoded on each group in group1 and group2, respectively.

  • fdr_target (float (default: 0.05)) – Tag features as DE based on posterior expected false discovery rate.

  • silent (bool (default: False)) – If True, disables the progress bar. Default: False.

  • weights (Optional[Literal['uniform', 'importance']] (default: 'uniform')) – Weights to use for sampling. If None, defaults to “uniform”.

  • filter_outlier_cells (bool (default: False)) – Whether to filter outlier cells with filter_outlier_cells().

  • importance_weighting_kwargs (dict | None (default: None)) – Keyword arguments passed into get_importance_weights().

  • dataloader (Iterator[dict[str, Tensor | None]] | None (default: None)) – Inference dataloader to materialize when the model was initialized without AnnData.

  • **kwargs – Keyword args for scvi.model.base.DifferentialComputation.get_bayes_factors()

Return type:

DataFrame

Returns:

Differential expression DataFrame.

DRVI.get_aggregated_posterior(adata=None, indices=None, batch_size=None, dof=3.0)[source]#

Compute the aggregated posterior over the u latent representations.

Parameters:

  • adata (default: None) – AnnData object to use. Defaults to the AnnData object used to initialize the model.

  • indices (default: None) – Indices of cells to use.

  • batch_size (default: None) – Batch size to use for computing the latent representation.

  • dof (default: 3.0) – Degrees of freedom for the Student’s t-distribution components. If None, components are Normal.

Returns:

A mixture distribution of the aggregated posterior.

DRVI.get_anndata_manager(adata, required=False)[source]#

Retrieves the AnnDataManager for a given AnnData object.

Requires self.id has been set. Checks for an AnnDataManager specific to this model instance.

Parameters:

  • adata (AnnData | MuData) – AnnData object to find a manager instance for.

  • required (bool (default: False)) – If True, errors on missing manager. Otherwise, returns None when manager is missing.

Return type:

AnnDataManager | None

DRVI.get_batch_representation(adata=None, indices=None, batch_size=None)[source]#

Get the batch representation for a given set of indices.

Return type:

ndarray

DRVI.get_effect_of_splits_out_of_distribution(embed, n_steps=20, n_samples=100, add_to_counts=1.0, directional=True, batch_size=128, seed=None)[source]#

Out-of-distribution per-split, per-gene effect scores via latent traversal.

Leverages DRVI’s additive decoder: each latent factor Z_i is decoded by its own subnetwork f_i, and the per-gene effects are aggregated over factors (logsumexp here). The effect of perturbing Z_i on a gene therefore depends only on f_i, which is probed by traversing the dimension over discrete steps between its observed min and max (from set_latent_dimension_stats()), averaged over randomly sampled batch/covariate combinations. f_i’s nonlinearity is summarized by its maximum contributions over the traversal.

For each factor and gene, three log-fold-change (LFC) effect scores are returned, each (n_splits, n_genes) (or (2, n_splits, n_genes) when directional, splitting the factor’s negative [min, 0] and positive [0, max] traversal directions):

  • max_possible – the largest achievable LFC: the factor’s extreme contribution vs. its own minimum, with every other factor held at its minimum contribution. Overall shows how much a gene can be moved by the factor, regardless of other factors.

  • min_possible – the same LFC but with every other factor held at its maximum contribution. Shrinks when a gene is driven by several factors. Overall specific is a factor’s effect.

  • combined – the product max_possible * min_possible; large only when the factor

is both strong and specific.

add_to_counts is a pseudo-count (relative to a 1e6-count reference cell) folded into LFC calculations for numerical stability. See the DRVI paper [Moinfar and Theis, 2024] (Supplementary Note “Interpretation of Latent Factors via Additive Decoder Architecture”) for the full derivation.

This function Requires embed.var to contain min and max.

The batch/covariate combinations are sampled randomly. Pass seed for a reproducible sample; None (default) leaves the RNG untouched, so it still honors a previously set scvi.settings.seed.

Return type:

dict[str, ndarray]

DRVI.get_effect_of_splits_within_distribution(adata=None, dataloader=None, add_to_counts=1.0, deterministic=True, directional=True, aggregations='ALL', skip_threshold=1.0, **kwargs)[source]#

In-distribution per-split, per-gene effect scores.

Aggregates each factor’s per-gene effect over the observed data, weighting cells by how strongly the factor is activated.

Parameters:

  • adata (AnnData | None (default: None)) – AnnData to run on. Defaults to the model’s registered AnnData.

  • dataloader (Iterator[dict[str, Tensor | None]] | None (default: None)) – Custom minibatch iterator used instead of building one from adata. Exactly one of adata / dataloader may be given.

  • add_to_counts (float (default: 1.0)) – Pseudo-count forwarded to the effect computation.

  • deterministic (bool (default: True)) – If True (default), use the posterior mean for the bottleneck (no sampling).

  • directional (bool (default: True)) – If True (default), score the factor’s positive and negative directions separately. Requires n_split_latent == n_latent.

  • aggregations (Sequence[Literal['max', 'linear_weighted_mean', 'exp_weighted_mean']] | str (default: 'ALL')) – Which cell-aggregation(s) to compute: "max" (peak effect across cells), "linear_weighted_mean" and "exp_weighted_mean" (activation-weighted means), or "ALL" (default) for all three. A single name or a sequence is accepted.

  • skip_threshold (float (default: 1.0)) – Cells whose latent activation is below this value contribute (near) zero weight to the weighted-mean aggregations, so weakly-activated cells are effectively ignored.

  • kwargs (Any) – Forwarded to iterate_on_effect_of_splits_within_distribution().

Return type:

dict[str, ndarray]

Returns:

A dict mapping each requested aggregation to a (n_splits, n_genes) array (or (2, n_splits, n_genes) when directional).

DRVI.get_elbo(adata=None, indices=None, batch_size=None, dataloader=None, return_mean=True, data_loader_kwargs=None, **kwargs)[source]#

Compute the evidence lower bound (ELBO) on the data.

The ELBO is the reconstruction error plus the Kullback-Leibler (KL) divergences between the variational distributions and the priors. It is different from the marginal log-likelihood; specifically, it is a lower bound on the marginal log-likelihood plus a term that is constant with respect to the variational distribution. It still gives good insights on the modeling of the data and is fast to compute.

Parameters:

  • adata (AnnData | None (default: None)) – AnnData object with var_names in the same order as the ones used to train the model. If None and dataloader is also None, it defaults to the object used to initialize the model.

  • indices (Sequence[int] | None (default: None)) – Indices of observations in adata to use. If None, defaults to all observations. Ignored if dataloader is not None.

  • batch_size (int | None (default: None)) – Minibatch size for the forward pass. If None, defaults to scvi.settings.batch_size. Ignored if dataloader is not None.

  • dataloader (Iterator[dict[str, Tensor | None]] | None (default: None)) – An iterator over minibatches of data on which to compute the metric. The minibatches should be formatted as a dictionary of Tensor with keys as expected by the model. If None, a dataloader is created from adata.

  • return_mean (bool (default: True)) – Whether to return the mean of the ELBO or the ELBO for each observation.

  • data_loader_kwargs (dict | None (default: None)) – Keyword args for data loader, in dict form.

  • **kwargs – Additional keyword arguments to pass into the forward method of the module.

Return type:

float

Returns:

Evidence lower bound (ELBO) of the data.

Notes

This is not the negative ELBO, so higher is better.

DRVI.get_feature_correlation_matrix(adata=None, indices=None, n_samples=10, batch_size=64, rna_size_factor=1000, transform_batch=None, correlation_type='spearman', silent=True)[source]#

Generate gene-gene correlation matrix using scvi uncertainty and expression.

Parameters:

  • adata (AnnData | None (default: None)) – AnnData object with equivalent structure to initial AnnData. If None, defaults to the AnnData object used to initialize the model.

  • indices (list[int] | None (default: None)) – Indices of cells in adata to use. If None, all cells are used.

  • n_samples (int (default: 10)) – Number of posterior samples to use for estimation.

  • batch_size (int (default: 64)) – Minibatch size for data loading into model. Defaults to scvi.settings.batch_size.

  • rna_size_factor (int (default: 1000)) – size factor for RNA prior to sampling gamma distribution.

  • transform_batch (list[int | float | str] | None (default: None)) –

    Batches to condition on. If transform_batch is:

    • None, then real observed batch is used.

    • int, then batch transform_batch is used.

    • list of int, then values are averaged over provided batches.

  • correlation_type (Literal['spearman', 'pearson'] (default: 'spearman')) – One of “pearson”, “spearman”.

  • %(de_silent)s

Return type:

DataFrame

Returns:

Gene-gene correlation matrix

DRVI.get_from_registry(adata, registry_key)[source]#

Returns the object in AnnData associated with the key in the data registry.

AnnData object should be registered with the model prior to calling this function via the self._validate_anndata method.

Parameters:

  • registry_key (str) – key of object to get from the data registry.

  • adata (AnnData | MuData) – AnnData to pull data from.

Return type:

ndarray

Returns:

The requested data as a NumPy array.

DRVI.get_importance_weights(adata, indices, qz, px, zs, max_cells=1024, truncation=False, n_mc_samples=500, n_mc_samples_per_pass=250, **data_loader_kwargs)[source]#

Computes importance weights for the given samples.

This method computes importance weights for every latent code in zs as a way to encourage latent codes providing high likelihoods across many cells in the considered subpopulation.

Parameters:

  • adata (AnnData | None) – Data to use for computing importance weights.

  • indices (list[int] | None) – Indices of cells in adata to use.

  • distributions – Dictionary of distributions associated with indices.

  • qz (Distribution) – Variational posterior distributions of the cells, aligned with indices.

  • px (Distribution) – Count distributions of the cells, aligned with indices.

  • zs (Tensor) – Samples associated with indices.

  • max_cells (int (default: 1024)) – Maximum number of cells used to estimated the importance weights

  • truncation (bool (default: False)) – Whether importance weights should be truncated. If True, the importance weights are truncated as described in [Ionides, 2008]. In particular, the provided value is used to threshold importance weights as a way to reduce the variance of the estimator.

  • n_mc_samples (int (default: 500)) – Number of Monte Carlo samples to use for estimating the importance weights, by default 500

  • n_mc_samples_per_pass (int (default: 250)) – Number of Monte Carlo samples to use for each pass, by default 250

  • **data_loader_kwargs – Keyword args for data loader.

Return type:

ndarray

Returns:

importance_weights Numpy array containing importance weights aligned with the provided indices.

Notes

This method assumes a normal prior on the latent space.

DRVI.get_interpretability_scores(embed, adata, key='OOD_combined', directional=True, gene_symbols=None, order_col='order', title_col='title', hide_vanished=True)[source]#

Return interpretability scores as a genes (rows) × dimensions (cols) DataFrame.

Reads scores stored in embed.varm by calculate_interpretability_scores() and the per-dimension annotations from set_latent_dimension_stats(), returning a tidy, ordered and labeled table for inspection.

Parameters:

  • embed (AnnData) – AnnData of the latent space with scores in varm and the set_latent_dimension_stats() annotations in var.

  • adata (AnnData) – Data AnnData, used to take the gene names for the DataFrame columns.

  • key (str (default: 'OOD_combined')) – varm score key to read (default "OOD_combined"). With directional, the _positive / _negative suffixes are appended automatically.

  • directional (bool (default: True)) – If True (default), include both the positive and negative direction of each factor as separate columns (titles suffixed with + / -).

  • gene_symbols (str | None (default: None)) – adata.var column to use for gene names; defaults to adata.var_names.

  • order_col (str (default: 'order')) – embed.var column used to order the dimension columns (default "order").

  • title_col (str (default: 'title')) – embed.var column used to label the dimension columns (default "title").

  • hide_vanished (bool (default: True)) – If True (default), drop factors (or directions) flagged as vanished by set_latent_dimension_stats().

Return type:

DataFrame

Returns:

A DataFrame of genes (rows) × dimensions (columns).

DRVI.get_latent_library_size(adata=None, indices=None, give_mean=True, batch_size=None, dataloader=None, **data_loader_kwargs)[source]#

Returns the latent library size for each cell.

This is denoted as \(\ell_n\) in the scVI paper.

Parameters:

  • adata (AnnData | None (default: None)) – AnnData object with equivalent structure to initial AnnData. If None, defaults to the AnnData object used to initialize the model.

  • indices (list[int] | None (default: None)) – Indices of cells in adata to use. If None, all cells are used.

  • give_mean (bool (default: True)) – Return the mean or a sample from the posterior distribution.

  • batch_size (int | None (default: None)) – Minibatch size for data loading into model. Defaults to scvi.settings.batch_size.

  • dataloader (Iterator[dict[str, Tensor | None]] | None (default: None)) – An iterator over minibatches of data on which to compute the metric. The minibatches should be formatted as a dictionary of Tensor with keys as expected by the model. If None, a dataloader is created from adata.

  • **data_loader_kwargs – Keyword args for data loader.

Return type:

ndarray

DRVI.get_latent_representation(adata=None, indices=None, give_mean=True, mc_samples=5000, batch_size=None, return_dist=False, dataloader=None, **data_loader_kwargs)[source]#

Compute the latent representation of the data.

This is typically denoted as \(z_n\).

Parameters:

  • adata (AnnData | None (default: None)) – AnnData object with var_names in the same order as the ones used to train the model. If None and dataloader is also None, it defaults to the object used to initialize the model.

  • indices (Sequence[int] | None (default: None)) – Indices of observations in adata to use. If None, defaults to all observations. Ignored if dataloader is not None

  • give_mean (bool (default: True)) – If True, returns the mean of the latent distribution. If False, returns an estimate of the mean using mc_samples Monte Carlo samples.

  • mc_samples (int (default: 5000)) – Number of Monte Carlo samples to use for the estimator for distributions with no closed-form mean (e.g., the logistic normal distribution). Not used if give_mean is True or if return_dist is True.

  • batch_size (int | None (default: None)) – Minibatch size for the forward pass. If None, defaults to scvi.settings.batch_size. Ignored if dataloader is not None

  • return_dist (bool (default: False)) – If True, returns the mean and variance of the latent distribution. Otherwise, returns the mean of the latent distribution.

  • dataloader (Iterator[dict[str, Tensor | None]] (default: None)) – An iterator over minibatches of data on which to compute the metric. The minibatches should be formatted as a dictionary of Tensor with keys as expected by the model. If None, a dataloader is created from adata.

  • **data_loader_kwargs – Keyword args for data loader.

Return type:

ndarray[tuple[Any, ...], dtype[TypeVar(_ScalarT, bound= generic)]] | tuple[ndarray[tuple[Any, ...], dtype[TypeVar(_ScalarT, bound= generic)]], ndarray[tuple[Any, ...], dtype[TypeVar(_ScalarT, bound= generic)]]]

Returns:

An array of shape (n_obs, n_latent) if return_dist is False. Otherwise, returns a tuple of arrays (n_obs, n_latent) with the mean and variance of the latent distribution.

DRVI.get_likelihood_parameters(adata=None, indices=None, n_samples=1, give_mean=False, batch_size=None, dataloader=None, **data_loader_kwargs)[source]#

Estimates for the parameters of the likelihood \(p(x \mid z)\).

Parameters:

  • adata (AnnData | None (default: None)) – AnnData object with equivalent structure to initial AnnData. If None, defaults to the AnnData object used to initialize the model.

  • indices (list[int] | None (default: None)) – Indices of cells in adata to use. If None, all cells are used.

  • n_samples (int | None (default: 1)) – Number of posterior samples to use for estimation.

  • give_mean (bool | None (default: False)) – Return expected value of parameters or a samples

  • batch_size (int | None (default: None)) – Minibatch size for data loading into model. Defaults to scvi.settings.batch_size.

  • dataloader (Iterator[dict[str, Tensor | None]] | None (default: None)) – An iterator over minibatches of data on which to compute the metric. The minibatches should be formatted as a dictionary of Tensor with keys as expected by the model. If None, a dataloader is created from adata.

  • **data_loader_kwargs – Keyword args for data loader.

Return type:

dict[str, ndarray]

DRVI.get_marginal_ll(adata=None, indices=None, n_mc_samples=1000, batch_size=None, return_mean=True, dataloader=None, data_loader_kwargs=None, **kwargs)[source]#

Compute the marginal log-likehood of the data.

The computation here is a biased estimator of the marginal log-likelihood of the data.

Parameters:

  • adata (AnnData | None (default: None)) – AnnData object with var_names in the same order as the ones used to train the model. If None and dataloader is also None, it defaults to the object used to initialize the model.

  • indices (Sequence[int] | None (default: None)) – Indices of observations in adata to use. If None, defaults to all observations. Ignored if dataloader is not None.

  • n_mc_samples (int (default: 1000)) – Number of Monte Carlo samples to use for the estimator. Passed into the module’s marginal_ll method.

  • batch_size (int | None (default: None)) – Minibatch size for the forward pass. If None, defaults to scvi.settings.batch_size. Ignored if dataloader is not None.

  • return_mean (bool (default: True)) – Whether to return the mean of the marginal log-likelihood or the marginal-log likelihood for each observation.

  • dataloader (Iterator[dict[str, Tensor | None]] (default: None)) – An iterator over minibatches of data on which to compute the metric. The minibatches should be formatted as a dictionary of Tensor with keys as expected by the model. If None, a dataloader is created from adata.

  • data_loader_kwargs (dict | None (default: None)) – Keyword args for data loader, in dict form.

  • **kwargs – Additional keyword arguments to pass into the module’s marginal_ll method.

Return type:

float | Tensor

Returns:

If True, returns the mean marginal log-likelihood. Otherwise returns a tensor of shape (n_obs,) with the marginal log-likelihood for each observation.

Notes

This is not the negative log-likelihood, so higher is better.

DRVI.get_normalized_expression(adata=None, indices=None, transform_batch=None, gene_list=None, library_size=1, n_samples=1, n_samples_overall=None, weights=None, batch_size=None, return_mean=True, return_numpy=None, silent=True, dataloader=None, data_loader_kwargs=None, **importance_weighting_kwargs)[source]#

Returns the normalized (decoded) gene expression.

This is denoted as \(\rho_n\) in the scVI paper.

Parameters:

  • adata (AnnData | None (default: None)) – AnnData object with equivalent structure to initial AnnData. If None, defaults to the AnnData object used to initialize the model.

  • indices (list[int] | None (default: None)) – Indices of cells in adata to use. If None, all cells are used.

  • transform_batch (list[int | float | str] | None (default: None)) – Batch to condition on. If transform_batch is: - None, then real observed batch is used. - int, then batch transform_batch is used. - Otherwise based on string

  • gene_list (list[str] | None (default: None)) – Return frequencies of expression for a subset of genes. This can save memory when working with large datasets and few genes are of interest.

  • library_size (Union[float, Literal['latent']] (default: 1)) – Scale the expression frequencies to a common library size. This allows gene expression levels to be interpreted on a common scale of relevant magnitude. If set to “latent”, use the latent library size.

  • n_samples (int (default: 1)) – Number of posterior samples to use for estimation.

  • n_samples_overall (int (default: None)) – Number of posterior samples to use for estimation. Overrides n_samples.

  • weights (Optional[Literal['uniform', 'importance']] (default: None)) – Weights to use for sampling. If None, defaults to “uniform”.

  • batch_size (int | None (default: None)) – Minibatch size for data loading into model. Defaults to scvi.settings.batch_size.

  • return_mean (bool (default: True)) – Whether to return the mean of the samples.

  • return_numpy (bool | None (default: None)) – Return a ndarray instead of a DataFrame. DataFrame includes gene names as columns. If either n_samples=1 or return_mean=True, defaults to False. Otherwise, it defaults to True.

  • %(de_silent)s

  • dataloader (Iterator[dict[str, Tensor | None]] | None (default: None)) – An iterator over minibatches of data on which to compute the metric. The minibatches should be formatted as a dictionary of Tensor with keys as expected by the model. If None, a dataloader is created from adata.

  • data_loader_kwargs (dict | None (default: None)) – Keyword args for data loader, in dict form.

  • importance_weighting_kwargs – Keyword arguments passed into get_importance_weights().

Return type:

ndarray | DataFrame

Returns:

If n_samples is provided and return_mean is False, this method returns a 3d tensor of shape (n_samples, n_cells, n_genes). If n_samples is provided and return_mean is True, it returns a 2d tensor of shape (n_cells, n_genes). In this case, return type is DataFrame unless return_numpy is True. Otherwise, the method expects n_samples_overall to be provided and returns a 2d tensor of shape (n_samples_overall, n_genes).

DRVI.get_reconstruction_effect_of_each_split(adata=None, dataloader=None, add_to_counts=1.0, aggregate_over_cells=True, deterministic=True, directional=False, **kwargs)[source]#

Effect of each split on reconstruction (summed over genes).

Used by set_latent_dimension_stats() to rank latent factors by how much they drive the reconstruction.

Parameters:

  • adata (AnnData | None (default: None)) – AnnData to run on. Defaults to the model’s registered AnnData.

  • dataloader (Iterator[dict[str, Tensor | None]] | None (default: None)) – Custom minibatch iterator used instead of building one from adata. Exactly one of adata / dataloader may be given.

  • add_to_counts (float (default: 1.0)) – Pseudo-count forwarded to the effect computation.

  • aggregate_over_cells (bool (default: True)) – If True (default), sum the effect over all cells; otherwise return per-cell values.

  • deterministic (bool (default: True)) – If True (default), use the posterior mean for the bottleneck (no sampling).

  • directional (bool (default: False)) – If True, split into the factor’s positive and negative directions.

  • kwargs (Any) – Forwarded to iterate_on_effect_of_splits_within_distribution().

Return type:

ndarray

Returns:

(n_splits,) (or (2, n_splits) if directional) when aggregating over cells, else (n_cells, n_splits) (or (n_cells, 2, n_splits)).

DRVI.get_reconstruction_error(adata=None, indices=None, batch_size=None, dataloader=None, return_mean=True, data_loader_kwargs=None, **kwargs)[source]#

Compute the reconstruction error on the data.

The reconstruction error is the negative log likelihood of the data given the latent variables. It is different from the marginal log-likelihood, but still gives good insights on the modeling of the data and is fast to compute. This is typically written as \(p(x \mid z)\), the likelihood term given one posterior sample.

Parameters:

  • adata (AnnData | None (default: None)) – AnnData object with var_names in the same order as the ones used to train the model. If None and dataloader is also None, it defaults to the object used to initialize the model.

  • indices (Sequence[int] | None (default: None)) – Indices of observations in adata to use. If None, defaults to all observations. Ignored if dataloader is not None

  • batch_size (int | None (default: None)) – Minibatch size for the forward pass. If None, defaults to scvi.settings.batch_size. Ignored if dataloader is not None

  • dataloader (Iterator[dict[str, Tensor | None]] | None (default: None)) – An iterator over minibatches of data on which to compute the metric. The minibatches should be formatted as a dictionary of Tensor with keys as expected by the model. If None, a dataloader is created from adata.

  • return_mean (bool (default: True)) – Whether to return the mean reconstruction loss or the reconstruction loss for each observation.

  • data_loader_kwargs (dict | None (default: None)) – Keyword args for data loader, in dict form.

  • **kwargs – Additional keyword arguments to pass into the forward method of the module.

Return type:

dict[str, float]

Returns:

Reconstruction error for the data.

Notes

This is not the negative reconstruction error, so higher is better.

DRVI.get_setup_arg(setup_arg)[source]#

Returns the string provided to setup of a specific setup_arg.

Return type:

attrdict

DRVI.get_state_registry(registry_key)[source]#

Returns the state registry for the AnnDataField registered with this instance.

Return type:

attrdict

DRVI.get_var_names(legacy_mudata_format=False)[source]#

Variable names of input data.

Return type:

dict

DRVI.iterate_on_ae_output(adata=None, dataloader=None, indices=None, batch_size=None, deterministic=False)[source]#

Yield (inference_outputs, generative_outputs) per minibatch.

Runs the full autoencoder (encoder + decoder) over adata (or a custom dataloader, e.g. from an out-of-core datamodule) via module.forward(compute_loss=False) (no loss computed). With deterministic=True the bottleneck uses the posterior mean.

DRVI.iterate_on_decoded_latent_samples(z, lib=None, batch_values=None, cat_values=None, cont_values=None, batch_size=128, map_cat_values=False)[source]#

Yield decoder (generative) outputs for given latent samples, batched.

Parameters:

  • z (ndarray) – Latent samples, shape (n_samples, n_latent).

  • lib (ndarray | None (default: None)) – Library sizes, shape (n_samples,). Defaults to 1e4 per sample.

  • batch_values (ndarray | None (default: None)) – Batch indices, shape (n_samples,). Defaults to 0.

  • cat_values (ndarray | None (default: None)) – Categorical covariates, shape (n_samples, n_cat_covs).

  • cont_values (ndarray | None (default: None)) – Continuous covariates, shape (n_samples, n_cont_covs).

  • batch_size (int (default: 128)) – Minibatch size.

  • map_cat_values (bool (default: False)) – Map categorical / batch values to integer codes via the AnnData manager registries.

DRVI.iterate_on_effect_of_splits_within_distribution(adata=None, dataloader=None, add_to_counts=1.0, deterministic=True, directional=True, **kwargs)[source]#

Yield (effect_tensor, latent) per minibatch: each split’s effect on every gene.

Low-level building block behind the in-distribution scores: it runs the autoencoder over the data and converts the per-split decoder parameters into a per-gene effect. The effect formula depends on module.split_aggregation ("logsumexp" or "mean").

Parameters:

  • adata (AnnData | None (default: None)) – AnnData to run on. Defaults to the model’s registered AnnData.

  • dataloader (Iterator[dict[str, Tensor | None]] | None (default: None)) – Custom minibatch iterator (e.g. from an out-of-core datamodule) used instead of building one from adata. Exactly one of adata / dataloader may be given.

  • add_to_counts (float (default: 1.0)) – Pseudo-count (relative to a 1e6-count cell) used by the "logsumexp" effect for numerical stability.

  • deterministic (bool (default: True)) – If True (default), use the posterior mean for the bottleneck (no sampling).

  • directional (bool (default: True)) – If True (default), split each effect into the factor’s positive and negative directions. Requires n_split_latent == n_latent.

  • kwargs (Any) – Forwarded to iterate_on_ae_output() (e.g. indices, batch_size).

Yields:

  • effect_tensor – Per-split, per-gene effect, (n_cells, n_splits, n_genes) (or (n_cells, 2, n_splits, n_genes) when directional).

  • latent – Posterior-mean latent for the minibatch, (n_cells, n_latent).

classmethod DRVI.load(dir_path, adata=None, accelerator='auto', device='auto', prefix=None, backup_url=None, datamodule=None, allowed_classes_names_list=None)[source]#

Instantiate a model from the saved output.

Parameters:

  • dir_path (str) – Path to saved outputs.

  • adata (AnnData | MuData | None (default: None)) – AnnData organized in the same way as data used to train model. It is not necessary to run setup_anndata, as AnnData is validated against the saved scvi setup dictionary. If None, will check for and load anndata saved with the model. If False, will load the model without AnnData.

  • accelerator (str (default: 'auto')) – Supports passing different accelerator types (“cpu”, “gpu”, “tpu”, “ipu”, “hpu”, “mps, “auto”) as well as custom accelerator instances.

  • device (int | str (default: 'auto')) – The device to use. Can be set to a non-negative index (int or str) or “auto” for automatic selection based on the chosen accelerator. If set to “auto” and accelerator is not determined to be “cpu”, then device will be set to the first available device.

  • prefix (str | None (default: None)) – Prefix of saved file names.

  • backup_url (str | None (default: None)) – URL to retrieve saved outputs from if not present on disk.

  • datamodule (LightningDataModule | None (default: None)) – EXPERIMENTAL A LightningDataModule instance to use for training in place of the default DataSplitter. Can only be passed in if the model was not initialized with AnnData.

  • allowed_classes_names_list (list[str] | None (default: None)) – list of allowed classes names to be loaded (besides the original class name)

Returns:

Model with loaded state dictionaries.

Examples

>>> model = ModelClass.load(save_path, adata)
>>> model.get_....
classmethod DRVI.load_query_data(adata=None, reference_model=None, registry=None, inplace_subset_query_vars=False, accelerator='auto', device='auto', unfrozen=False, freeze_dropout=False, freeze_expression=True, freeze_decoder_first_layer=True, freeze_batchnorm_encoder=True, freeze_batchnorm_decoder=False, freeze_classifier=True, transfer_batch=True, datamodule=None)[source]#

Online update of a reference model with scArches algorithm [Lotfollahi et al., 2021].

Parameters:

  • adata (AnnData | MuData (default: None)) – AnnData organized in the same way as data used to train model. It is not necessary to run setup_anndata, as AnnData is validated against the registry.

  • reference_model (str | BaseModelClass (default: None)) – Either an already instantiated model of the same class or a path to saved outputs for the reference model.

  • inplace_subset_query_vars (bool (default: False)) – Whether to subset and rearrange query vars inplace based on vars used to train the reference model.

  • accelerator (str (default: 'auto')) – Supports passing different accelerator types (“cpu”, “gpu”, “tpu”, “ipu”, “hpu”, “mps, “auto”) as well as custom accelerator instances.

  • device (int | str (default: 'auto')) – The device to use. Can be set to a non-negative index (int or str) or “auto” for automatic selection based on the chosen accelerator. If set to “auto” and accelerator is not determined to be “cpu”, then device will be set to the first available device.

  • unfrozen (bool (default: False)) – Override all other freeze options for a fully unfrozen model

  • freeze_dropout (bool (default: False)) – Whether to freeze dropout during training

  • freeze_expression (bool (default: True)) – Freeze neurons corresponding to expression in first layer

  • freeze_decoder_first_layer (bool (default: True)) – Freeze neurons corresponding to first layer in decoder

  • freeze_batchnorm_encoder (bool (default: True)) – Whether to freeze batchnorm weight and bias during training for encoder

  • freeze_batchnorm_decoder (bool (default: False)) – Whether to freeze batchnorm weight and bias during training for decoder

  • freeze_classifier (bool (default: True)) – Whether to freeze classifier completely. Only applies to SCANVI.

  • transfer_batch (bool (default: True)) – Allow for surgery on the batch covariate. Only applies to SYSVI.

  • datamodule (LightningDataModule | None (default: None)) – EXPERIMENTAL A LightningDataModule instance to use for training in place of the default DataSplitter. Can only be passed in if the model was not initialized with AnnData.

static DRVI.load_registry(dir_path, prefix=None)[source]#

Return the full registry saved with the model.

Parameters:

  • dir_path (str) – Path to saved outputs.

  • prefix (str | None (default: None)) – Prefix of saved file names.

Return type:

dict

Returns:

The full registry saved with the model

DRVI.minify_adata(minified_data_type='latent_posterior_parameters', use_latent_qzm_key='X_latent_qzm', use_latent_qzv_key='X_latent_qzv')[source]#

Minify the model’s adata.

Minifies the AnnData object associated with the model according to the method specified by minified_data_type and registers the new fields with the model’s AnnDataManager. This also sets the minified_data_type attribute of the underlying BaseModuleClass instance.

Parameters:

  • minified_data_type (Literal['latent_posterior_parameters'] (default: 'latent_posterior_parameters')) –

    Method for minifying the data. One of the following:

    • "latent_posterior_parameters": Store the latent posterior mean and variance in

      obsm using the keys use_latent_qzm_key and use_latent_qzv_key.

    • "latent_posterior_parameters_with_counts": Store the latent posterior mean and

      variance in obsm using the keys use_latent_qzm_key and use_latent_qzv_key, and the raw count data in X.

  • use_latent_qzm_key (str (default: 'X_latent_qzm')) – Key to use for storing the latent posterior mean in obsm when minified_data_type is "latent_posterior".

  • use_latent_qzv_key (str (default: 'X_latent_qzv')) – Key to use for storing the latent posterior variance in obsm when minified_data_type is "latent_posterior".

Return type:

None

Notes

The modification is not done inplace – instead the model is assigned a new (minified) version of the AnnData.

DRVI.posterior_predictive_sample(adata=None, indices=None, transform_batch=None, n_samples=1, gene_list=None, batch_size=None, dataloader=None, silent=True, **data_loader_kwargs)[source]#

Generate predictive samples from the posterior predictive distribution.

The posterior predictive distribution is denoted as \(p(\hat{x} \mid x)\), where \(x\) is the input data and \(\hat{x}\) is the sampled data.

We sample from this distribution by first sampling n_samples times from the posterior distribution \(q(z \mid x)\) for a given observation, and then sampling from the likelihood \(p(\hat{x} \mid z)\) for each of these.

Parameters:

  • adata (AnnData | None (default: None)) – AnnData object with an equivalent structure to the model’s dataset. If None, defaults to the AnnData object used to initialize the model.

  • indices (list[int] | None (default: None)) – Indices of the observations in adata to use. If None, defaults to all the observations.

  • transform_batch (list[int | float | str] | None (default: None)) – Batch to condition on. If transform_batch is: - None, then real observed batch is used. - int, then batch transform_batch is used. - Otherwise based on string

  • n_samples (int (default: 1)) – Number of Monte Carlo samples to draw from the posterior predictive distribution for each observation.

  • gene_list (list[str] | None (default: None)) – Names of the genes to which to subset. If None, defaults to all genes.

  • batch_size (int | None (default: None)) – Minibatch size to use for data loading and model inference. Defaults to scvi.settings.batch_size. Passed into BaseModelClass._make_data_loader.

  • dataloader (Iterator[dict[str, Tensor | None]] | None (default: None)) – An iterator over minibatches of data on which to compute the metric. The minibatches should be formatted as a dictionary of Tensor with keys as expected by the model. If None, a dataloader is created from adata.

  • **data_loader_kwargs – Keyword args for data loader.

Return type:

GCXS

Returns:

Sparse multidimensional array of shape (n_obs, n_vars) if n_samples == 1, else (n_obs, n_vars, n_samples).

static DRVI.prepare_query_anndata(adata, reference_model, return_reference_var_names=False, inplace=True)[source]#

Prepare data for query integration.

This function will return a new AnnData object with padded zeros for missing features, as well as correctly sorted features.

Parameters:

  • adata (AnnData) – AnnData organized in the same way as data used to train model. It is not necessary to run setup_anndata, as AnnData is validated against the registry.

  • reference_model (str | BaseModelClass) – Either an already instantiated model of the same class or a path to saved outputs for the reference model.

  • return_reference_var_names (bool (default: False)) – Only load and return reference var names if True.

  • inplace (bool (default: True)) – Whether to subset and rearrange query vars inplace or return new AnnData.

Return type:

AnnData | Index | None

Returns:

Query adata ready to use in load_query_data unless return_reference_var_names in which case a pd.Index of reference var names is returned.

static DRVI.prepare_query_mudata(mdata, reference_model, return_reference_var_names=False, inplace=True)[source]#

Prepare multimodal dataset for query integration.

This function will return a new MuData object such that the AnnData objects for individual modalities are given padded zeros for missing features, as well as correctly sorted features.

Parameters:

  • mdata (MuData) – MuData organized in the same way as data used to train the model. It is not necessary to run setup_mudata, as MuData is validated against the registry.

  • reference_model (str | BaseModelClass) – Either an already instantiated model of the same class or a path to saved outputs for the reference model.

  • return_reference_var_names (bool (default: False)) – Only load and return reference var names if True.

  • inplace (bool (default: True)) – Whether to subset and rearrange query vars inplace or return new MuData.

Return type:

MuData | dict[str, Index] | None

Returns:

Query mudata ready to use in load_query_data unless return_reference_var_names in which case a dictionary of pd.Index of reference var names is returned.

classmethod DRVI.register_manager(adata_manager)[source]#

Registers an AnnDataManager instance with this model class.

Stores the AnnDataManager reference in a class-specific manager store. Intended for use in the setup_anndata() class method followed up by retrieval of the AnnDataManager via the _get_most_recent_anndata_manager() method in the model init method.

Notes

Subsequent calls to this method with an AnnDataManager instance referring to the same underlying AnnData object will overwrite the reference to previous AnnDataManager.

DRVI.save(dir_path, prefix=None, overwrite=False, save_anndata=False, save_kwargs=None, legacy_mudata_format=False, datamodule=None, **anndata_write_kwargs)[source]#

Save the state of the model.

Neither the trainer optimizer state nor the trainer history are saved. Model files are not expected to be reproducibly saved and loaded across versions until we reach version 1.0.

Parameters:

  • dir_path (str) – Path to a directory.

  • prefix (str | None (default: None)) – Prefix to prepend to saved file names.

  • overwrite (bool (default: False)) – Overwrite existing data or not. If False and directory already exists at dir_path, an error will be raised.

  • save_anndata (bool (default: False)) – If True, also saves the anndata

  • save_kwargs (dict | None (default: None)) – Keyword arguments passed into save().

  • legacy_mudata_format (bool (default: False)) – If True, saves the model var_names in the legacy format if the model was trained with a MuData object. The legacy format is a flat array with variable names across all modalities concatenated, while the new format is a dictionary with keys corresponding to the modality names and values corresponding to the variable names for each modality.

  • datamodule (LightningDataModule | None (default: None)) – EXPERIMENTAL A LightningDataModule instance to use for training in place of the default DataSplitter. Can only be passed in if the model was not initialized with AnnData.

  • anndata_write_kwargs – Kwargs for write()

DRVI.set_latent_dimension_stats(embed, adata=None, dataloader=None, vanished_threshold=0.5)[source]#

Annotate embed.var with per-dimension reconstruction effect, ordering and stats.

Typically the first interpretability step: it ranks the latent factors and records the per-dimension activation stats. When a split groups several latent dimensions (n_split_latent < n_latent), each split’s reconstruction effect is broadcast to the latent dimensions it contains.

Parameters:

  • embed (AnnData) – AnnData of the latent space (one var per latent factor, X holding the latent values). Annotated in place.

  • adata (AnnData | None (default: None)) – AnnData to compute the reconstruction effect on. Defaults to the model’s registered AnnData.

  • dataloader (Iterator[dict[str, Tensor | None]] | None (default: None)) – Custom minibatch iterator used instead of building one from adata. Exactly one of adata / dataloader may be given.

  • vanished_threshold (float (default: 0.5)) – A dimension (or direction) is flagged as “vanished” (inactive) when its maximum absolute latent value stays below this threshold.

Return type:

None

Notes

Adds the columns reconstruction_effect, order, max_value, mean, min, max, std, std_abs, title and vanished (+ per-direction vanished flags) to embed.var.

classmethod DRVI.setup_annbatch(cls, collection_path=None, paths=None, batch_key=None, labels_key=None, sample_key=None, unlabeled_category='Unknown', layer=None, categorical_covariate_keys=None, continuous_covariate_keys=None, rebuild=True, batch_size=4096, chunk_size=256, preload_nchunks=32, preload_to_gpu=True, dataset_size='20GB', shuffle=False, var_subset=None, merge=None, adatas=None, use_class_sampler=False, class_sampler_key=None, class_weights=None)[source]#

Set up a model from on-disk AnnData files via the annbatch streaming loader.

Builds (or reuses) a zarr-backed DatasetCollection from the supplied h5ad file paths, then wraps it in a AnnbatchDataModule ready for training.

Parameters:

  • collection_path (str | None (default: None)) – Directory where the zarr collection is written (or already exists). If None (default), a path is auto-generated as "./{ModelName}_annbatch.zarr" in the current working directory.

  • paths (list[str] | None (default: None)) – Paths to h5ad files that make up the training dataset. If None, an existing collection at collection_path is opened without rebuilding — useful when the zarr store was created by a previous call. Must be provided when the store does not exist yet.

  • batch_key (str | None (default: None)) – Column in obs to use as the batch variable.

  • labels_key (str | None (default: None)) – Column in obs to use as the cell-type / label variable.

  • sample_key (str | None (default: None)) – Column in obs to use as the sample variable. Used by models like MrVI.

  • unlabeled_category (str (default: 'Unknown')) – Value used to mark unlabeled cells in labels_key. Required by semi-supervised models such as SCANVI. Defaults to "Unknown".

  • layer (str | None (default: None)) – Layer in the h5ad files to use as the count matrix. None uses adata.X (falling back to adata.raw.X when a raw slot exists).

  • categorical_covariate_keys (list[str] | None (default: None)) – Additional categorical covariate columns in obs.

  • continuous_covariate_keys (list[str] | None (default: None)) – Additional continuous covariate columns in obs.

  • rebuild (bool (default: True)) – If True, always rebuild the zarr collection even if it already exists on disk. If False (default), reuse an existing collection and skip the (potentially expensive) add_adatas step.

  • batch_size (int (default: 4096)) – Number of cells per batch yielded by the Loader.

  • chunk_size (int (default: 256)) – Number of cells loaded from disk contiguously per read.

  • preload_nchunks (int (default: 32)) – Number of chunks to preload and shuffle in memory.

  • preload_to_gpu (bool (default: True)) – Whether the loader should move data to GPU before yielding.

  • dataset_size (int | str (default: '20GB')) – Number of observations to load into memory for shuffling / pre-processing when building the collection, or annbatch’s human-readable size strings, e.g. "20GB".

  • shuffle (bool (default: False)) – Whether to pre-shuffle cells when building the collection.

  • var_subset (list[str] | None (default: None)) – Optional list of gene names to restrict the collection to (passed as var_subset to add_adatas).

  • merge (Optional[Literal['same', 'unique', 'first', 'only']] (default: None)) – How annbatch should merge var metadata across inputs. Passed through to annbatch.DatasetCollection.add_adatas().

Returns:

AnnbatchDataModule A configured datamodule whose registry can be passed directly to the model constructor, e.g. model = SCVI(registry=dm.registry).

Notes

After training, saving and reloading the model requires reconstructing the datamodule manually and passing it to load() via the datamodule argument — the same pattern used by all custom datamodules.

classmethod DRVI.setup_anndata(adata, layer=None, batch_key=None, labels_key=None, size_factor_key=None, categorical_covariate_keys=None, continuous_covariate_keys=None, **kwargs)[source]#

Sets up the AnnData object for this model.

A mapping will be created between data fields used by this model to their respective locations in adata. None of the data in adata are modified. Only adds fields to adata.

Parameters:

  • adata (AnnData) – AnnData object. Rows represent cells, columns represent features.

  • layer (str | None (default: None)) – if not None, uses this as the key in adata.layers for raw count data.

  • batch_key (str | None (default: None)) – key in adata.obs for batch information. Categories will automatically be converted into integer categories and saved to adata.obs[‘_scvi_batch’]. If None, assigns the same batch to all the data.

  • labels_key (str | None (default: None)) – key in adata.obs for label information. Categories will automatically be converted into integer categories and saved to adata.obs[‘_scvi_labels’]. If None, assigns the same label to all the data.

  • size_factor_key (str | None (default: None)) – key in adata.obs for size factor information. Instead of using library size as a size factor, the provided size factor column will be used as offset in the mean of the likelihood. Assumed to be on linear scale.

  • categorical_covariate_keys (list[str] | None (default: None)) – keys in adata.obs that correspond to categorical data. These covariates can be added in addition to the batch covariate and are also treated as nuisance factors (i.e., the model tries to minimize their effects on the latent space). Thus, these should not be used for biologically-relevant factors that you do _not_ want to correct for.

  • continuous_covariate_keys (list[str] | None (default: None)) – keys in adata.obs that correspond to continuous data. These covariates can be added in addition to the batch covariate and are also treated as nuisance factors (i.e., the model tries to minimize their effects on the latent space). Thus, these should not be used for biologically-relevant factors that you do _not_ want to correct for.

DRVI.to_device(device)[source]#

Move the model to the device.

Parameters:

device (str | int | device) – Device to move model to. Options: ‘cpu’ for CPU, integer GPU index (e.g., 0), ‘cuda:X’ where X is the GPU index (e.g. ‘cuda:0’), or a torch.device object (including XLA devices for TPU). See torch.device for more info.

Examples

>>> adata = scvi.data.synthetic_iid()
>>> model = scvi.model.SCVI(adata)
>>> model.to_device("cpu")  # moves model to CPU
>>> model.to_device("cuda:0")  # moves model to GPU 0
>>> model.to_device(0)  # also moves model to GPU 0
DRVI.train(max_epochs=None, accelerator='auto', devices='auto', train_size=None, validation_size=None, shuffle_set_split=True, load_sparse_tensor=False, batch_size=128, early_stopping=False, datasplitter_kwargs=None, plan_config=None, plan_kwargs=None, datamodule=None, trainer_config=None, **trainer_kwargs)[source]#

Train the model.

Parameters:

  • max_epochs (int | None (default: None)) – The maximum number of epochs to train the model. The actual number of epochs may be less if early stopping is enabled. If None, defaults to a heuristic based on get_max_epochs_heuristic(). Must be passed in if datamodule is passed in, and it does not have an n_obs attribute.

  • accelerator (str (default: 'auto')) – Supports passing different accelerator types (“cpu”, “gpu”, “tpu”, “ipu”, “hpu”, “mps, “auto”) as well as custom accelerator instances.

  • devices (int | list[int] | str (default: 'auto')) – The devices to use. Can be set to a non-negative index (int or str), a sequence of device indices (list or comma-separated str), the value -1 to indicate all available devices, or “auto” for automatic selection based on the chosen accelerator. If set to “auto” and accelerator is not determined to be “cpu”, then devices will be set to the first available device.

  • train_size (float | None (default: None)) – Float, or None. Size of training set in the range [0.0, 1.0]. The default is None, which is practically 0.9 and potentially adding a small last batch to validation cells. Passed into DataSplitter. Not used if datamodule is passed in.

  • validation_size (float | None (default: None)) – Size of the test set. If None, defaults to 1 - train_size. If train_size + validation_size < 1, the remaining cells belong to a test set. Passed into DataSplitter. Not used if datamodule is passed in.

  • shuffle_set_split (bool (default: True)) – Whether to shuffle indices before splitting. If False, the val, train, and test set are split in the sequential order of the data according to validation_size and train_size percentages. Passed into DataSplitter. Not used if datamodule is passed in.

  • load_sparse_tensor (bool (default: False)) – EXPERIMENTAL If True, loads data with sparse CSR or CSC layout as a Tensor with the same layout. Can lead to speedups in data transfers to GPUs, depending on the sparsity of the data. Passed into DataSplitter. Not used if datamodule is passed in.

  • batch_size (int (default: 128)) – Minibatch size to use during training. Passed into DataSplitter. Not used if datamodule is passed in.

  • early_stopping (bool (default: False)) – Perform early stopping. Additional arguments can be passed in through **kwargs. See Trainer for further options.

  • datasplitter_kwargs (dict | None (default: None)) – Additional keyword arguments passed into DataSplitter. Values in this argument can be overwritten by arguments directly passed into this method, when appropriate. Not used if datamodule is passed in.

  • plan_config (Mapping[str, Any] | KwargsConfig | None (default: None)) – Configuration object or mapping used to build TrainingPlan. Values in plan_kwargs and explicit arguments take precedence.

  • plan_kwargs (Mapping[str, Any] | KwargsConfig | None (default: None)) – Additional keyword arguments passed into TrainingPlan. Values in this argument can be overwritten by arguments directly passed into this method, when appropriate.

  • datamodule (LightningDataModule | None (default: None)) – EXPERIMENTAL A LightningDataModule instance to use for training in place of the default DataSplitter. Can only be passed in if the model was not initialized with AnnData.

  • trainer_config (Mapping[str, Any] | KwargsConfig | None (default: None)) – Configuration object or mapping used to build Trainer. Values in trainer_kwargs and explicit arguments take precedence.

  • **kwargs – Additional keyword arguments passed into Trainer.

DRVI.transfer_fields(adata, **kwargs)[source]#

Transfer fields from a model to an AnnData object.

Return type:

AnnData

DRVI.update_setup_method_args(setup_method_args)[source]#

Update setup method args.

Parameters:

setup_method_args (dict) – This is a bit of a misnomer, this is a dict representing kwargs of the setup method that will be used to update the existing values in the registry of this instance.

DRVI.view_anndata_setup(adata=None, hide_state_registries=False)[source]#

Print summary of the setup for the initial AnnData or a given AnnData object.

Parameters:

  • adata (AnnData | MuData | None (default: None)) – AnnData object setup with setup_anndata or transfer_fields().

  • hide_state_registries (bool (default: False)) – If True, prints a shortened summary without details of each state registry.

Return type:

None

DRVI.view_registry(hide_state_registries=False)[source]#

Prints summary of the registry.

Parameters:

hide_state_registries (bool (default: False)) – If True, prints a shortened summary without details of each state registry.

Return type:

None

static DRVI.view_setup_args(dir_path, prefix=None)[source]#

Print args used to setup a saved model.

Parameters:

  • dir_path (str) – Path to saved outputs.

  • prefix (str | None (default: None)) – Prefix of saved file names.

Return type:

None

DRVI.view_setup_method_args()[source]#

Prints setup kwargs used to produce a given registry.

Parameters:

registry – Registry produced by an AnnDataManager.

Return type:

None

Contents