Parameter introduction

stSCI provides two primary functions for its core workflow:

stSCI.train(): Train a new stSCI model from your single-cell and spatial transcriptomics data.
stSCI.eval(): Load a pre-trained stSCI model and perform inference.

This section provides a detailed explanation of the parameters for each of these two functions.

stSCI.train

def train(
    sc_adata: sc.AnnData,
    st_adata: sc.AnnData,
    multi_slice_key: Optional[str] = None,
    hvg_count: int = 3000,
    batch_sim_k: int = 25,
    overall_sim_k: int = 100,
    cluster_key: str = 'cluster',
    init_trans_matrix: bool = False,
    lr: float = 1e-3,
    epochs: int = 500,
    update_iter: int = 50,
    model_dims: List[int] = [512, 30],
    clustering: bool = False,
    cluster_method: str = 'mclust',
    cluster_para: Union[int, float] = 0.8,
    deconvolution: bool = False,
    coor_reconstruction: bool = False,
    model_save_path: Optional[str] = None,
    device: str = 'cuda' if (torch.cuda.is_available()) else 'cpu'
) -> Tuple[sc.AnnData]:

sc_adata: A sc.AnnData object containing the single-cell transcriptomics data.
st_adata: A sc.AnnData object containing the spatial transcriptomics data.
multi_slice_key: A str specifying the key in st_adata.obs that indicates different ST batches. This is used for analysising multi-slice spatial datasets, set to None if you have only a single slice.
hvg_count: An int specifying the number of highly variable genes (HVGs) to select for model training.
batch_sim_k: An int specifying the number of most similar single cells to sample for each spatial spot during the mini-batch training phase.
batch_sim_k: An int specifying the number of most similar single cells to consider for each spatial spot when inferring the final SC-ST relationships after the model is trained.
cluster_key: A str specifying the key in sc_adata.obs that contains the cell type labels for the single-cell data.
lr: A float specifying the learning rate for the model optimizer during training.
epochs: An int for the total number of training epochs to run.
update_iter: An int specifying the frequency (epochs) to update the MNN pairs used for data integration.
model_dims: AList[int] defining the architecture of the neural network. Each integer represents the number of neurons in a hidden layer.
clustering: A bool flag. If True, performs spatial domain clustering.
cluster_method: A str indicating the algorithm to use for clustering, we support mclust and louvain. This is only used if clustering is True.
cluster_para: A numeric input specifying the parameter for the chosen cluster_method. For Louvain, this is the resolution parameter. For mclust, this is the cluster number.
deconvolution: A bool flag. If True, performs deconvolution.
coor_reconstruction: A bool flag. If True, performs coordinate reconstruction.
model_save_path: A str providing the file path where the trained model parameters will be saved. If None, the model is not saved to disk.
device: A str specifying the computational device for model training (e.g., ‘cpu’, ‘cuda’, or ‘cuda:0’).

NOTE: stSCI fully supports training on a CPU, ensuring accessibility for users who do not have a dedicated GPU. However, due to the computational intensity of deep learning, using a CUDA-enabled GPU is strongly recommended. If you do not have a compatible GPU, please specify ‘cpu’ for parameter device.

stSCI.eval

def eval(
    sc_adata: sc.AnnData,
    st_adata: sc.AnnData,
    model_path: Union[torch.nn.Module, str],
    multi_slice_key: Optional[str] = None,
    hvg_count: int = 3000,
    overall_sim_k: int = 100,
    cluster_key: str = 'cluster',
    clustering: bool = False,
    cluster_method: str = 'mclust',
    cluster_para: Union[int, float] = 0.8,
    deconvolution: bool = False,
    coor_reconstruction: bool = False,
    imputation: bool = False
) -> Tuple[sc.AnnData]:

stSCI.eval shares most of its parameters with stSCI.train. The key difference is the model_path parameter, which is used to specify the pre-trained model for inference. It flexibly accepts either a str type data pointing to the saved model file or a pre-loaded torch.nn.Module object directly.