scREBOUND

scREBOUND is a pretrained single-cell foundation model that generate cell embedding for multi-purpose downstream tasks.

Dependency

python >= 3.12
anndata >= 0.11.4
sklearn >= 1.5.2
numpy >= 2.0.1
scanpy >= 1.11.1
pytorch == 2.5.0+cu124
flash-attn == 2.7.4.post1

Directory

• ./eval_adata/ stores the evaluation anndata, the download link
• ./model_statedict/ stores the trained scREBOUND model, the download link
• ./script_preprocess/ stores the preprocessing scripts for the training data
• ./script_train/ stores the training scripts for scREBOUND
• ./src/ stores the source code of scREBOUND
• ./script_infer.py is the cell embedding inference script using the pretrained foundation model

Usage

python script_infer.py --config [config_file.yaml]

The config_file.yaml stores the configuration of the inference, which include the key parameters:

• seed: the random seed
• device: running device
• model_dir: the directory of the model state dict
• data_dir: the directory of the evaluation dataset
• output_dir: the output directory See config_pancreas.yaml for the example setting on pancreas evaluation dataset.