spcoral.model.regist_model

spcoral.model.regist_model#

class spcoral.model.regist_model(adata_omics1, adata_omics2, graph_method, k_spatial_omics1=None, radius_spatial_omics1=None, k_spatial_omics2=None, radius_spatial_omics2=None, use_obsm='spatial', n_layer=[1, 2, 3, 4, 5], alpha=0.1, device=device(type='cuda', index=0), random_seed=2024, strict_repro=False, learning_rate=0.001, weight_decay=0.0001, epochs=100, gradient_clipping=5.0, hidden_dim_shared=32, out_dim_shared=8, hidden_dim_pcc=32, out_dim_pcc=8, GAN_batch_d_per_iter=5)#

Bases: object

A PyTorch-based model for cross-modality spatial omics integration via graph attention networks and adversarial alignment.

Parameters:

  • adata_omics1 (anndata.AnnData) – First omics AnnData object. Must contain obsm['feat'] (modality-specific features) and spatial coordinates (via use_obsm).

  • adata_omics2 (anndata.AnnData) – Second omics AnnData object with the same requirements.

  • graph_method (str) – Method used to construct spatial graphs (passed to adata_to_dgl; e.g., ‘knn’ or ‘radius’).

  • k_spatial_omics1 (int, optional) – Number of nearest neighbors for omics1 graph when graph_method='knn'.

  • radius_spatial_omics1 (float, optional) – Radius for omics1 graph when graph_method='radius'.

  • k_spatial_omics2 (int, optional) – Number of nearest neighbors for omics2 graph.

  • radius_spatial_omics2 (float, optional) – Radius for omics2 graph.

  • use_obsm (str, optional (default: 'spatial')) – Key in .obsm containing spatial coordinates.

  • n_layer (list of int, optional (default: [1, 2, 3, 4, 5])) – Orders of neighborhood aggregation (hop distances) for morphological feature computation. The actual layers used are [0] + n_layer.

  • alpha (float, optional (default: 0.1)) – Weight balancing reconstruction loss and adversarial (GAN) loss.

  • device (torch.device, optional (default: torch.device('cuda:0'))) – Device for training.

  • random_seed (int, optional (default: 2024)) – Random seed for reproducibility.

  • strict_repro (bool, optional (default: False)) – If True, enforces stricter reproducibility (e.g., deterministic CUDA operations).

  • learning_rate (float, optional (default: 0.001)) – Learning rate for the main model optimizer.

  • weight_decay (float, optional (default: 0.0001)) – Weight decay for both optimizers.

  • epochs (int, optional (default: 100)) – Number of training epochs.

  • gradient_clipping (float, optional (default: 5.0)) – Maximum gradient norm for clipping.

  • hidden_dim_shared (int, optional (default: 32)) – Hidden dimension in the shared feature branch.

  • out_dim_shared (int, optional (default: 8)) – Output dimension of the shared embedding.

  • hidden_dim_pcc (int, optional (default: 32)) – Hidden dimension in modality-specific (PCC) branches.

  • out_dim_pcc (int, optional (default: 8)) – Output dimension of modality-specific embeddings.

  • GAN_batch_d_per_iter (int, optional (default: 5)) – Number of discriminator updates per generator update in the Wasserstein GAN.

adata_omics1, adata_omics2

Processed AnnData objects with added embeddings after training.

Type:

anndata.AnnData

model#

The main COM_NET encoder-decoder model (instantiated during train).

Type:

torch.nn.Module

__init__(adata_omics1, adata_omics2, graph_method, k_spatial_omics1=None, radius_spatial_omics1=None, k_spatial_omics2=None, radius_spatial_omics2=None, use_obsm='spatial', n_layer=[1, 2, 3, 4, 5], alpha=0.1, device=device(type='cuda', index=0), random_seed=2024, strict_repro=False, learning_rate=0.001, weight_decay=0.0001, epochs=100, gradient_clipping=5.0, hidden_dim_shared=32, out_dim_shared=8, hidden_dim_pcc=32, out_dim_pcc=8, GAN_batch_d_per_iter=5)#
Parameters:

  • adata_omics1 (AnnData) –

  • adata_omics2 (AnnData) –

  • graph_method (str) –

  • k_spatial_omics1 (int | None) –

  • radius_spatial_omics1 (float | None) –

  • k_spatial_omics2 (int | None) –

  • radius_spatial_omics2 (float | None) –

  • use_obsm (str) –

  • n_layer (List[int]) –

  • alpha (float) –

  • device (device) –

  • random_seed (int) –

  • strict_repro (bool) –

  • learning_rate (float) –

  • weight_decay (float) –

  • epochs (int) –

  • gradient_clipping (float) –

  • hidden_dim_shared (int) –

  • out_dim_shared (int) –

  • hidden_dim_pcc (int) –

  • out_dim_pcc (int) –

  • GAN_batch_d_per_iter (int) –

Methods

__init__(adata_omics1, adata_omics2, ...[, ...])

train()

Train the integration model.
train()#

Train the integration model.

Returns:

  • adata_omics1 : anndata.AnnData Updated first omics object with obsm['embedding'] and obsm['share_feature'].

  • adata_omics2 : anndata.AnnData Updated second omics object with the same added keys.

  • loss_list : list of [reconstruction_loss, gan_loss] per epoch Training loss history.

Return type:

tuple

Contents