This repository hosts the official implementation of GEARS, a method that can predict transcriptional response to both single and multi-gene perturbations using single-cell RNA-sequencing data from perturbational screens.

Install PyG, and then do pip install cell-gears.

• Fixed training breakpoint bug from v0.1.0
• Preprocessed dataloader now available for Replogle 2022 RPE1 and K562 essential datasets
• Added custom split, fixed no-test split

• GEARS is currently not designed to handle training across multiple cell types or cross-cell type transfer of predictions
• GEARS has not been tested for training with bulk sequencing data.
• When trained on single-gene perturbation data alone, GEARS cannot reliably predict outcomes for combinatorial perturbations. The model must be trained on some combinatorial perturbation data to make such predictions.
• GEARS has been tested using datasets that contain multiple perturbation types, and multiple cells within each perturbation type. Datasets with too few cells per perturbation or too few perturbations may not work well with our model.

Using the API, you can (1) reproduce the results in our paper and (2) train GEARS on your perturbation dataset using a few lines of code.

from gears import PertData, GEARS

# get data
pert_data = PertData('./data')
# load dataset in paper: norman, adamson, dixit.
pert_data.load(data_name = 'norman')
# specify data split
pert_data.prepare_split(split = 'simulation', seed = 1)
# get dataloader with batch size
pert_data.get_dataloader(batch_size = 32, test_batch_size = 128)

# set up and train a model
gears_model = GEARS(pert_data, device = 'cuda:8')
gears_model.model_initialize(hidden_size = 64)
gears_model.train(epochs = 20)

# save/load model
gears_model.save_model('gears')
gears_model.load_pretrained('gears')

# predict
gears_model.predict([['CBL', 'CNN1'], ['FEV']])
gears_model.GI_predict(['CBL', 'CNN1'], GI_genes_file=None)

To use your own dataset, create a scanpy adata object with a gene_name column in adata.var, and two columns condition, cell_type in adata.obs. Then run:

pert_data.new_data_process(dataset_name = 'XXX', adata = adata)
# to load the processed data
pert_data.load(data_path = './data/XXX')

@article{roohani2023predicting,
  title={Predicting transcriptional outcomes of novel multigene perturbations with gears},
  author={Roohani, Yusuf and Huang, Kexin and Leskovec, Jure},
  journal={Nature Biotechnology},
  year={2023},
  publisher={Nature Publishing Group US New York}
}

Paper: Link

Code for reproducing figures: Link