ERGO is a deep learing based model for predicting TCR-peptide binding.

pytorch 1.0.0
numpy 1.15.4
scikit-learn 0.19.2

• The code has been tested on Linux 3.10.0-957.27.2.el7.x86_64.
• The code runs faster with GPU usage, but it should work also on CPU. It was developed on GPU (Nvidia Tesla K40m with CUDA 10.0).

For initialization, download this repository or clone using git clone https://github.com/IdoSpringer/ERGO. It should take a few seconds.

Training a model should take about an hour. All runs use the main ERGO.py file.

For training the ERGO model, run:

python ERGO.py train model_type dataset sampling device 

When: model_type is ae for the TCR-autoencoder based model, or lstm for the lstm based model. dataset is mcpas for McPAS-TCR dataset, or vdjdb for VDJdb dataset. sampling can be specific for distinguishing different binders, naive for separating non-binders and binders, or memory for distinguishing binders and memory TCRs. device is a CUDA GPU device (e.g. 'cuda:0') or 'cpu' for CPU device.

Add the flag --protein suit the model for protein binding instead of specific peptide binding. Add the argument --train_auc_file=file to write down the model train AUC during training. Add the argument --test_auc_file=file to write down the model validation AUC during training. Add the argument --model_file=file.pt in order to save the trained model. Add the argument --test_data_file=file.pickle in order to save the test data.

The autoencoder based model requires a pre-trained TCR-autoencoder. for training the TCR-autoencoder, go to the TCR-Autoencoder directory using

cd TCR_Autoencoder

and run:

python train_tcr_autoencoder.py BM_data_CDR3s device model_file.pt

when device is a CUDA GPU device (e.g. 'cuda:0') or 'cpu' for CPU device. The trained autoencoder will be saved in model_file as a pytorch model.

When you run the ERGO.py file, use the argument --ae_file=trained_autoencoder_file. You can use the already trained tcr_autoencoder.pt model instead, with --ae_file=auto.

In order to evaluate the model for specific peptides or proteins, run:

python ERGO.py test model_type dataset sampling device --model_file=file.pt --test_data_file=file.pickle

When: model_type is ae for the TCR-autoencoder based model, or lstm for the lstm based model. dataset is mcpas for McPAS-TCR dataset, or vdjdb for VDJdb dataset. sampling can be specific for distinguishing different binders, naive for separating non-binders and binders, or memory for distinguishing binders and memory TCRs. device is a CUDA GPU device (e.g. 'cuda:0') or 'cpu' for CPU device. (All arguments same as the model which is now evaluated)

The argument --model_file=file.pt is the trained model to be evaluated. The argument --test_data_file=file.pickle is the test data (which the model has not seen during training, in order to do so please save it in the training run command).

Add the flag --protein suit the model for protein binding instead of specific peptide binding.

You can use the models you have trained to predict new data, or you can use our already trained models in the models directory.

Data for prediction must be in a .csv format, where the first column is the TCRs and the second column is the peptides. See example.

For binding prediction, run:

python ERGO.py predict model_type dataset sampling device --model_file=file.pt --test_data_file=file.csv

When: model_type is ae for the TCR-autoencoder based model, or lstm for the lstm based model. dataset is mcpas for McPAS-TCR dataset, or vdjdb for VDJdb dataset. sampling can be specific for distinguishing different binders, naive for separating non-binders and binders, or memory for distinguishing binders and memory TCRs. device is a CUDA GPU device (e.g. 'cuda:0') or 'cpu' for CPU device.

The argument --model_file=file.pt is the trained model. If you want to use our trained models, use --model_file=auto, and the code will choose the right trained model according to the arguments model_type, dataset and sampling. If you are using your model file, make sure that those arguments match the model.

The argument --test_data_file=file.csv is the test data to predict in .csv format. If you want to run our example use --test_data_file=auto.

The code should print the input TCRs and the peptides, with the predicted binding probabilities. Notice that in the autoencoder model, TCR max length is 28, so longer sequences will be ignored. Prediction should be take a few seconds.

1. Springer, I., Besser, H., Tickotsky-Moskovitz, N., Dvorkin, S. & Louzoun, Y. Prediction of specific TCR-peptide binding from large dictionaries of TCR-peptide pairs. bioRxiv 650861 (2019). doi:10.1101/650861