STAPLER

Preprint (bioRxiv) . Report Bug

1. Clone the repo

   git clone https://github.com/NKI-AI/STAPLER.git

2. Navigate to the directory containing setup.py

   cd STAPLER

3. Install the STAPLER package (should take less than 10 minutes)

    python -m pip install .

The following data is available here:

• data/train:
  • TCR and peptide datasets used to pre-train STAPLER
  • The training dataset
  • 5 train and validation folds created from the training dataset used to fine-tune STAPLER
• data/test:
  • VDJDB+ and VDJDB+ETN test datasets used to test STAPLER
• model/pretrained_model:
  • 1 pre-trained model checkpoint
• model/finetuned_model:
  • 5 fine-tuned model checkpoints (one for each fold).
• predictions:
  • 5 predictions for each fold on the VDJDB+ETN test set
  • 1 ensembled prediction of all 5-fold predictions on the VDJDB+ETN test set

STAPLER was pre-trained and fine-tuned using an a100 GPU. At this moment no other GPU's have been tested.

Inside the tools directory the following file should be changed:

• .env.example: Which is an environment file with paths to data, model checkpoints and output paths. It should be adapt the .env.example to your local file-system and then change the file-name to .env. For more information see python-dotenv.

Inside the tools directory contains the following files to pre-train, fine-tune and/or test STAPLER on a SLURM cluster. Also provide an argument to --partition to specify the partition to use.

• sbatch pretrain_STAPLER.sh: Pre-train STAPLER.
• sbatch train_STAPLER.sh: Fine-tune STAPLER using 5-fold cross-validation.
• sbatch test_STAPLER.sh: Test on a test set using a fine-tuned model checkpoint.

Alternatively run STAPLER directly on a machine with an appropriate GPU (see requirements).

• python pretrain.py: Pre-train STAPLER.
• python train_5_fold.py: Fine-tune STAPLER using 5-fold cross-validation.
• python test.py: Test on a test set using a fine-tuned model checkpoint.

The pre-training should take a day, fine-tuning should take a couple of hours per fold and testing/inference should take a couple of minutes for all 5-fold predictions.

To experiment with custom model parameters change the paramteres inside the config directory (implemented using Hydra). The config directory contains the following main configuration files:

• pretrain.yaml: Configuration parameters file for pre-training.
• train_5_fold.yaml: Configuration parameters file for fine-tuning.
• test.yaml: Configuration parameters file for testing.

To request a feature or to discuss any issues, please let us know by opening an issue on the issues page.

• The notebooks used to make the pre-print figures will be available soon

Corresponding author: Ton Schumacher

Ton Schumacher group (NKI) - Group website -

Ai for Oncology group (NKI) - Group website -

The development of the STAPLER model is the result of a collaboration between the Schumacher lab AIforOncology lab at the Netherlands Cancer Institute. The following people contributed to the development of the model:

• Bjørn Kwee (implementation, development, evaluation, refactoring)
• Marius Messemaker (supervision, development, evaluation)
• Eric Marcus (supervision, refactoring)
• Wouter Scheper (supervision)
• Jonas Teuwen (supervision)
• Ton Schumacher (supervision)

A part of the data was provided, and consequent results were interpreted by the following people from the Wu lab (DFCI and Harvard Medical School):

• Giacomo Oliveira
• Catherine Wu

STAPLER is built on top of the x-transformers package

Distributed under the Apache 2.0 License.

(back to top)