EDDISPLICE

EDDISPLICE is a single-modality splicing-only VAE that uses the same missingness-aware partial encoder as SpliceVI but without a gene expression branch. It is useful as a baseline or when only splicing data is available.

---

When to use EDDISPLICE vs SpliceVI

	EDDISPLICE	SpliceVI
Input	Splicing only	Gene expression + Splicing
Use case	Splicing-only datasets; ablation baselines	Paired or unpaired multimodal data
Latent space	Splicing-driven	Joint GE + AS

---

API reference

splicevi.EDDISPLICE

Bases: VAEMixin, UnsupervisedTrainingMixin, BaseModelClass

__init__(adata=None, code_dim=16, h_hidden_dim=64, encoder_hidden_dim=128, encoder_n_layers=2, latent_dim=10, dropout_rate=0.1, learn_concentration=True, splice_likelihood='dirichlet_multinomial', encode_covariates=False, deeply_inject_covariates=True, initialize_embeddings_from_pca=True, pool_mode='mean', max_nobs=-1, latent_distribution='normal', **kwargs)

Initialize an EDDISPLICE model wrapping a PARTIALVAE with the PartialEncoderEDDIFaster encoder and a linear decoder.

This model learns a low-dimensional latent representation of junction-level splicing usage from an AnnData object set up via :meth:EDDISPLICE.setup_anndata. The encoder uses per-junction embeddings plus an EDDI-style partial observation mechanism that only aggregates observed junctions per cell. The reconstruction head supports binomial, beta-binomial, or Dirichlet–multinomial likelihoods.

Parameters:

Name	Type	Description	Default
adata	AnnData | None	AnnData object with layers/fields registered using :meth:EDDISPLICE.setup_anndata. If None, the module is initialized lazily at training time.	None
code_dim	int	Dimensionality of the per-junction embedding table used by the encoder.	16
h_hidden_dim	int	Hidden size of the shared "h" MLP that processes each observed junction embedding + PSI value before pooling across junctions.	64
encoder_hidden_dim	int	Hidden size of the final encoder MLP that maps the pooled per-cell representation to the mean and variance of the latent Gaussian.	128
encoder_n_layers	int	Number of hidden layers in the encoder MLP (passed to :class:scvi.nn.FCLayers).	2
latent_dim	int	Dimensionality of the latent space :math:z.	10
dropout_rate	float	Dropout probability applied in the encoder networks.	0.1
learn_concentration	bool	If True, learn a global concentration parameter used in the beta-binomial or Dirichlet–multinomial reconstruction likelihood.	True
splice_likelihood	Literal['binomial', 'beta_binomial', 'dirichlet_multinomial']	Reconstruction likelihood for junction counts. One of: "binomial": simple binomial over junction vs. cluster counts. "beta_binomial": beta-binomial with learned concentration. "dirichlet_multinomial": hierarchical Dirichlet–multinomial over ATSE-level totals using a junction→ATSE mapping.	'dirichlet_multinomial'
encode_covariates	bool	If True, concatenates observed covariates (batch and any extra covariates registered via setup_anndata) into the encoder input.	False
deeply_inject_covariates	bool	If True, passes covariates into the decoder so that reconstruction can be conditioned on batch / covariates as well as :math:z.	True
initialize_embeddings_from_pca	bool	If True and adata is provided, initialize the per-junction embedding table with a truncated SVD (PCA) fit to the registered splicing layer instead of random initialization.	True
pool_mode	Literal['mean', 'sum']	How to aggregate per-junction representations into a per-cell vector: "mean": average over observed junctions per cell. "sum": sum over observed junctions per cell.	'mean'
max_nobs	int	Maximum number of observed (cell, junction) pairs processed in a single chunk inside the encoder. If negative, disables chunking. Useful for controlling memory when masks are dense.	-1
latent_distribution	Literal['normal', 'ln']	Form of the latent prior/approximate posterior. "normal" uses a standard Gaussian; "ln" applies a softmax transformation to the latent samples to produce simplex-valued representations.	'normal'
**kwargs		Additional keyword arguments forwarded to :class:PARTIALVAE. These are rarely needed in routine use and are kept for extensibility.	{}

Notes

• For splice_likelihood="dirichlet_multinomial", a junction→ATSE mapping is constructed from adata.var["event_id"] and stored as a sparse tensor module.junc2atse.
• Per-junction embeddings can optionally be initialized from PCA for faster and more stable training on high-dimensional splicing matrices.

get_latent_representation(adata=None, indices=None, give_mean=True, batch_size=None)

Return latent embeddings of splicing VAE.

Parameters:

Name	Type	Description	Default
adata	AnnData | None	AnnData for inference (defaults to init adata).	None
indices	Sequence[int] | None	Cell indices to use.	None
give_mean	bool	If True, use posterior mean; else sample.	True
batch_size	int | None	Batch size.	None

Returns:

Type	Description
Array of shape (cells, latent_dim).

get_normalized_splicing(adata=None, indices=None, use_z_mean=True, n_samples=1, batch_size=None, return_numpy=False, silent=True)

Return the decoded splicing probabilities p_nj = sigmoid(decoder_logits).

Parameters:

Name	Type	Description	Default
adata	AnnData | None	AnnData for inference (defaults to the one used at init).	None
indices	Sequence[int] | None	Which cells to pull (default: all).	None
use_z_mean	bool	If True, run generative with use_z_mean=True.	True
n_samples	int	How many posterior samples to draw (passed to inference).	1
batch_size	int | None	Mini-batch size (defaults to scvi.settings.batch_size).	None
return_numpy	bool	If True, returns a (n_cells, n_junctions) numpy array; otherwise returns a DataFrame with var_names as columns.	False
silent	bool	If False, shows a little progress info.	True

Returns:

Type	Description
Array or DataFrame of shape (cells, junctions) of decoded probabilities.

setup_anndata(adata, junc_ratio_layer, junc_counts_layer, cluster_counts_layer, psi_mask_layer, batch_key=None, size_factor_key=None, categorical_covariate_keys=None, continuous_covariate_keys=None, **kwargs) classmethod

Set up AnnData for EddiSplice.

Parameters:

Name	Type	Description	Default
adata	AnnData	AnnData to register.	required
junc_ratio_layer	str	Layer with junction usage ratios (X input).	required
junc_counts_layer	str	Layer with junction counts (successes).	required
cluster_counts_layer	str	Layer with total cluster counts (trials).	required
psi_mask_layer	str	Layer with binary mask (1=observed, 0=missing) per junction.	required
batch_key	str | None	Column in obs for batch.	None
size_factor_key	str | None	If provided, registers size factor but unused for splicing.	None
categorical_covariate_keys	list[str] | None		None
continuous_covariate_keys	list[str] | None		None

train(max_epochs=200, lr=0.0001, accelerator='auto', devices='auto', train_size=None, validation_size=None, shuffle_set_split=True, batch_size=512, weight_decay=0.001, eps=1e-08, early_stopping=True, save_best=True, check_val_every_n_epoch=None, n_steps_kl_warmup=None, n_epochs_kl_warmup=10, reduce_lr_on_plateau=False, lr_factor=0.6, lr_patience=30, lr_threshold=0.0, lr_min=0.0, datasplitter_kwargs=None, plan_kwargs=None, **kwargs)

Trains the model using amortized variational inference on splicing data.

Parameters:

Name	Type	Description	Default
max_epochs	int	Number of epochs to train over.	200
lr	float	Learning rate for optimization.	0.0001
accelerator	str	Hardware acceleration options.	'auto'
devices	str	Hardware acceleration options.	'auto'
train_size	float | None	Proportions for splitting the data.	None
validation_size	float | None	Proportions for splitting the data.	None
shuffle_set_split	bool	Whether to shuffle indices before splitting.	True
batch_size	int	Minibatch size for training.	512
weight_decay	float	Optimizer hyperparameters.	0.001
eps	float	Optimizer hyperparameters.	0.001
early_stopping	bool	Early stopping and checkpointing options.	True
save_best	bool	Early stopping and checkpointing options.	True
check_val_every_n_epoch	int | None	Frequency of validation checks.	None
n_steps_kl_warmup	int | None	KL warmup parameters.	None
n_epochs_kl_warmup	int | None	KL warmup parameters.	None
datasplitter_kwargs	dict | None	Additional options for data splitting, training plan, and trainer.	None
plan_kwargs	dict | None	Additional options for data splitting, training plan, and trainer.	None
**kwargs	dict | None	Additional options for data splitting, training plan, and trainer.	None