scglue.models.scglue.SCGLUEModel

class scglue.models.scglue.SCGLUEModel(adatas, vertices, latent_dim=50, h_depth=2, h_dim=256, dropout=0.2, disc_norm=False, shared_batches=False, random_seed=0)[source]

Bases: Model

GLUE model for single-cell multi-omics data integration

Parameters:

  • adatas (typing.Mapping[str, AnnData]) – Datasets (indexed by modality name)

  • vertices (typing.List[str]) – Guidance graph vertices (must cover feature names in all modalities)

  • latent_dim (int) – Latent dimensionality

  • h_depth (int) – Hidden layer depth for encoder and discriminator

  • h_dim (int) – Hidden layer dimensionality for encoder and discriminator

  • dropout (float) – Dropout rate

  • disc_norm (bool) – Whether to perform normalization at discriminator input

  • shared_batches (bool) – Whether the same batches are shared across modalities

  • random_seed (int) – Random seed

Methods

adopt_pretrained_model

Adopt buffers and parameters from a pretrained model

classify_data

Obtain cell type classification

compile

Prepare model for training

decode_data

Decode data

encode_data

Compute data (cell) embedding

encode_graph

Compute graph (feature) embedding

fit

Fit model on given datasets

freeze_cells

Freeze cell embeddings

get_losses

Compute loss function values

unfreeze_cells

Unfreeze cell embeddings

upgrade

Upgrade the model if generated by older versions

Attributes

ALIGN_BURNIN_PRG

GRAPH_BATCHES

MAX_EPOCHS_PRG

PATIENCE_PRG

REDUCE_LR_PATIENCE_PRG

logger