sc_reconstruction.models.ReconPretrainedStateModel#

class sc_reconstruction.models.ReconPretrainedStateModel(checkpoint_path, protein_embeds_path, emb_key='X_state', read_depth=4.0, encode_batch_size=64, decode_batch_size=64, **kwargs)[source]#

Bases: BaseReconstructionModel

Parameters:

  • checkpoint_path (str)

  • protein_embeds_path (str)

  • emb_key (str)

  • read_depth (float)

  • encode_batch_size (int)

  • decode_batch_size (int)

__init__(checkpoint_path, protein_embeds_path, emb_key='X_state', read_depth=4.0, encode_batch_size=64, decode_batch_size=64, **kwargs)[source]#

Pre-trained State Embedding (SE) foundation model [Adduri et al., 2025].

Wraps the SE encoder for ReconEval’s foundation-model reconstruction task. Requires the state package and runs in the cstm_scvi_env conda env.

Args:

checkpoint_path: Path to the pre-trained model checkpoint protein_embeds_path: Path to protein embeddings emb_key: Key for cell embeddings in adata.obsm read_depth: Read depth for decoding batch_size: Batch size for inference library_size_mode: For interface compatibility (not used in this model)

Parameters:

  • checkpoint_path (str)

  • protein_embeds_path (str)

  • emb_key (str)

  • read_depth (float)

  • encode_batch_size (int)

  • decode_batch_size (int)

Methods

__init__(checkpoint_path, protein_embeds_path)

Pre-trained State Embedding (SE) foundation model [Adduri et al., 2025].

forward(x)

Forward pass for compatibility

get_latent_representation(X)

get_overlap_genes(genes)

Get overlapping genes between adata and protein embeddings.

load(path[, map_location])

Load model configuration

predict(X[, target_genes, read_depth])

Predict reconstruction using pre-trained model

predict_relu(X)

Predict reconstruction with ReLU activation

prepare([adata])

Cache the adata reference and its gene list on the wrapper.

save(path)

Save model configuration (not the actual pre-trained weights)

set_genes(genes)

Set the genes for reconstruction

train([datamodule])

No-op: SE is used frozen.