TITAN - Tcr epITope bimodal Attention Networks

The library itself has few dependencies (see setup.py) with loose requirements.

Create a virtual environment and install dependencies

python -m venv --system-site-packages venv
source venv/bin/activate
pip install -r requirements.txt

Install in editable mode for development:

pip install -e .

For data handling, we make use of the pytoda package. If you bring your own data, it needs to adhere to the following format:

• tcrs.csv A .csv file containing two columns, one for the tcr sequences and one for their IDs.
• epitopes.csv A .csv file containing two columns, one for the epitope sequences and one for their IDs. This can optionally also be a .smi file (tab-separated) with the SMILES seuqences of the eptiopes.
• train.csv A .csv file containing three columns, one for TCR IDs, one for epitope IDs and one for the labels. This data is used for training.
• test.csv A .csv file containing three columns, one for TCR IDs, one for epitope IDs and one for the labels. This data is used for testing.

NOTE: tcrs.csv and epitopes.csv need to contain all TCRs and epitopes used during training and testing. No duplicates in both sequence and IDs are allowed. All data can be found in https://ibm.box.com/v/titan-dataset .

The TITAN model uses the architecture published as 'paccmann_predictor' package. Example parameter files are given in the params folder.

python3 scripts/flexible_training.py \
name_of_training_data_files.csv \
name_of_testing_data_files.csv \
path_to_tcr_file.csv \
path_to_epitope_file.csv/.smi \
path_to_store_trained_model \
path_to_parameter_file \
training_name \
bimodal_mca

To load a TITAN model after pretraining and finetune it on another dataset, use the semifrozen_finetuning.py script. Use the parameter number_of_tunable_layers to control the number of layers which will be tuned, the rest will be frozen. Model will freeze epitope input channel first and the final dense layers last. Do not change the input data type (i.e. SMILES or amino acids) between pretraining and finetuning.

python3 scripts/semifrozen_finetuning.py \
name_of_training_data_files.csv \
name_of_testing_data_files.csv \
path_to_tcr_file.csv \
path_to_epitope_file.smi \
path_to_pretrained_model \
path_to_store_model \
training_name \
path_to_parameter_file \
bimodal_mca

A trained model is provided in https://ibm.box.com/v/titan-dataset along with the dataset used to train it. The model is pretrained on BindingDB and finetuned using the semifrozen setting, on full TCR sequences and with SMILES encoding of epitopes. All parameters can be found in the parameter files provided.

python3 scripts/flexible_model_eval.py \
name_of_test_data_file.csv \
path_to_tcr_file.csv \
path_to_epitope_file.smi \
path_to_trained_model_folder \
bimodal_mca \
save_name

The script scripts/knn_cv.py uses the KNN baseline model of the paper and performs a cross validation. The script can be used in two modes, shared and separate. Shared is the default mode as specified above. In separate mode, the TCRs and epitope sequences for training and testing dont need to be in the same file, but can be split across two files. To use this mode, simply provide additional paths to -test_tcr and test_ep arguments.

python3 scripts/knn_cv.py \
-d path_to_data_folder \
-tr name_of_training_data_files.csv \
-te name_of_testing_data_files.csv \
-f 10 \
-ep path_to_epitope_file.ccsv \
-tcr path_to_tcr_file.csv \
-r path_to_result_folder \
-k 25

type python3 scripts/knn_cv.py -h for help. The data in data_folder needs to be structured as:

data_path
├── fold0
│   ├── name_of_training_data_files.csv
│   ├── name_of_testing_data_files.csv
...
├── fold9
│   ├── name_of_training_data_files.csv
│   ├── name_of_testing_data_files.csv

To generate full sequences of TCRs from CDR3 sequence and V and J segment names, the cdr3_to_full_seq.py script can be used. The script relies on the user having downloaded a fasta files containing the Names of V and J segments with their respecive sequences called V_segment_sequences.fasta and J_segment_sequences.fasta. These can be downloaded from IMGT.org. Header names must be provided to the script to adapt to different format of the input file.

python3 scripts/cdr3_to_full_seq.py \
directoy_with_VJ_segment_fasta_files \
path_to_file_with_input_sequences.csv \
v_seq_header \
j_seq_header \
cdr3_header \
path_to_output_file.csv

If you use titan in your projects, please cite the following:

@article{weber2021titan
    author = {Weber, Anna and Born, Jannis and Rodriguez Martinez, Maria},
    title = "{TITAN: T-cell receptor specificity prediction with bimodal attention networks}",
    journal = {Bioinformatics},
    volume = {37},
    number = {Supplement_1},
    pages = {i237-i244},
    year = {2021},
    month = {07},
    issn = {1367-4803},
    doi = {10.1093/bioinformatics/btab294},
    url = {https://doi.org/10.1093/bioinformatics/btab294}
}