model

class model.CANAL_model(gpu_option='0', CLASS=7, EMBEDDING_DIM=64, VALIDATE_EVERY=1, PATIENCE=10, BATCH_SIZE=14, EPOCHS=100, LEARNING_RATE=0.0002)

Bases: object

Construct the CANAL model, which continually adapts pre-trained language model to universal annotation of scRNA-seq data

evaluation(pred_cell_type, true_celltype, novel_celltype=None)

If the test dataset has ground-truth cell-type labels, evaluate the performance of the CANAL model. If no novel cells in the test data, we will exhibit total annotation accuracy, F1 score as well as ARI; if novel cells exist in the test data, we will exhibit H-score, total annotation accuracy, annotation accuracy of known cells as well as accuracy of unknown cells

Pred_cell_type (array):

predicted cell types of the test dataset via the CANAL model

True_cell_type (array):

true cell types of the test dataset

Novel_celltype (array):

novel cell type of the test dataset that have never appeared in the fine-tuning data stream. The default value is None.

predict(adata_predict, ckpt_dir, experiments, stage_num, dataset, novel=False, temperature=0.8)

Predict cell types of an unlabelled dataset using the CANAL model, which has completed fine-tuning after certain stage

Adata_predict (AnnData object):

unlabeled test data we need to annotate

Ckpt_dir:

path of the saved fine-tuned CANAL model

Experiments (str):

name of experiments

Stage_num (int):

which stage of fine-tuning the CANAL model has completed

Dataset (str):

name of the dataset used during the given stage

Novel (bool):

whether to detect novel cells in the test data, default value is False. If True, we will annotate cells as “Unassigned”, if their uncertainty score is larger than the automatic determined threshold.

Temperature (float):

the temperature parameter in the energy function if novel cell detection is needed. The default value is 0.8

Returns:

an array, which provides predicted cell types of the test dataset

train(experiments, pre_dataset, dataset, adata, cell_type, current_stage, is_final_stage, ckpt_dir, rehearsal_size=1000, highly_variable_idx=None, lambda_KD=0.1, SEED=1)

Train the CANAL model at a certain stage

Experiments (str):

name of experiments

Pre_dataset (str):

name of previous dataset

Dataset (str):

name of current dataset

Adata (AnnData object):

current data used to fine-tune the model

Cell_type (array):

the corresponding cell types of current data

Current_stage(int):

which stage the model is fine-tuned

Stage_num (int):

current stage of fine-tuning

Is_final_stage (bool):

whether current stage is the final stage or not

Ckpt_dir:

path to save the model

Rehearsal_size (int):

the number of cells in total we can preserve in the example bank, default = 1000

Highly_variable_idx (array):

index of highly variable genes, should be provided if this is the initial stage

Lambda_KD (float):

trainin weight for the representation knowledge distillation loss, default = 0.1

Gpu_option (int):

which device to run the model, default = ‘0’

SEED (int):

which random seed is chosen to run the current model, default = 1

class model.Identity(SEQ_LEN, dropout=0.0, h_dim=64, out_dim=10)

Bases: Module

Construct network output layer

forward(x)

Forward propagation

training: bool

class model.SCDataset(data, label, CLASS)

Bases: Dataset

Construct SCDataset from AnnData

model.example_bank_update(example_bank_previous, adata, new_model_annotation, embedding, prototype, current_label_dict, current_label_set, current_label, each_class_num, current_stage)

Update current example bank via prototypes

model.get_embedding(data, CLASS, model)

get samples’ representation of the penultimate layer

model.save_model(experiments, dataset, stage_num, is_final_stage, example_bank, label_dict, highly_variable_idx, model, ckpt_dir)

Save current model, cell-type library as well as example bank for continual learning and evaluation.

Experiments (str):

name of experiments

Stage_num (int):

current stage of fine-tuning

Is_final_stage (bool):

whether current stage is the final stage or not;

Example_bank (AnnData object):

the selected examples for repeatedly reviewing

Label_dict (array):

cell type library with all the cell types that have appeared so far

Highly_variable_idx (array):

index of highly variable genes

Model:

the fine-tuned CANAL model after this stage

Ckpt_dir:

path to save the model

model.setup_seed(seed)

Choose a random seed to run the current model.

Seed (int):

random seed