omicverse.single.pyTOSICA

omicverse.single.pyTOSICA#

class omicverse.single.pyTOSICA(adata, project_path, gmt_path=None, label_name='Celltype', mask_ratio=0.015, max_g=300, max_gs=300, n_unannotated=1, embed_dim=48, depth=1, num_heads=4, batch_size=8, device='cuda:0', auto_download=True)[source]#

TOSICA wrapper for pathway-informed transformer-based cell-type annotation.

Parameters:

  • adata (anndata.AnnData) – Training/reference AnnData with labels.

  • project_path (str) – Output directory for TOSICA checkpoints and logs.

  • gmt_path (str|None, optional, default=None) – Pathway GMT file path. If None, default gene-set resources are used.

  • label_name (str, optional, default='Celltype') – Label column in adata.obs.

  • mask_ratio (float, optional, default=0.015) – Ratio of masked genes/tokens used for training regularization.

  • max_g (int, optional, default=300) – Maximum number of genes used per pathway/tokenization unit.

  • max_gs (int, optional, default=300) – Maximum number of gene sets used in the model.

  • n_unannotated (int, optional, default=1) – Number of unlabeled classes reserved during training.

  • embed_dim (int, optional, default=48) – Transformer embedding dimension.

  • depth (int, optional, default=1) – Number of transformer encoder layers.

  • num_heads (int, optional, default=4) – Number of attention heads.

  • batch_size (int, optional, default=8) – Mini-batch size used during training/inference.

  • device (str, optional, default='cuda:0') – Device used for model training/inference.

Returns:

Initializes TOSICA model configuration and training resources.

Return type:

None

Examples

>>> tosica_obj = ov.single.pyTOSICA(adata=ref_adata, project_path="./tosica")
Parameters:

auto_download (bool (default: True))

__init__(adata, project_path, gmt_path=None, label_name='Celltype', mask_ratio=0.015, max_g=300, max_gs=300, n_unannotated=1, embed_dim=48, depth=1, num_heads=4, batch_size=8, device='cuda:0', auto_download=True)[source]#

Initialize a pyTOSICA object for cell type classification.

Parameters:

  • adata (anndata.AnnData) – Training/reference AnnData for TOSICA.

  • project_path (str) – Directory used to save masks, labels, checkpoints, and logs.

  • gmt_path (str or None) – Pathway GMT identifier/path. If None, full-connection mask is used.

  • label_name (str) – Label column in adata.obs.

  • mask_ratio (float) – Random mask ratio used when pathway mask is unavailable.

  • max_g (int) – Maximum number of genes per pathway.

  • max_gs (int) – Maximum number of pathway tokens used by the model.

  • n_unannotated (int) – Number of extra unannotated tokens appended to pathway mask.

  • embed_dim (int) – Transformer embedding dimension.

  • depth (int) – Number of transformer layers.

  • num_heads (int) – Number of attention heads.

  • batch_size (int) – Training batch size.

  • device (str) – Preferred device string (for example 'cuda:0').

  • auto_download (bool (default: True))

Methods

__init__(adata, project_path[, gmt_path, ...])

Initialize a pyTOSICA object for cell type classification.

load([load_path])

Load a pre-trained TOSICA model.

predicted(pre_adata[, laten, n_step, ...])

Predict cell types for new single-cell data.

save([save_path])

Save the trained TOSICA model.

train([pre_weights, lr, epochs, lrf])

Train the TOSICA model for cell type classification.
Contents