piercelab/tcrmodel2

TCRmodel2

To model TCR-pMHC complex structures, as well as unbound TCR structures, with high fidelity.

While you have the option to download and install TCRmodel2 locally, we recommend utilizing our web server for generating predictions. The web server offers a user-friendly interface and eliminates the need for local installation. You can access the web server at the following URL:

https://tcrmodel.ibbr.umd.edu/

If you use our tool, please cite:

Yin R, Ribeiro-Filho HV, Lin V, Gowthaman R, Cheung M, Pierce BG. (2023) TCRmodel2: high-resolution modeling of T cell receptor recognition using deep learning. Nucleic Acids Research, 51(W1):W569-W576. https://doi.org/10.1093/nar/gkad356

Quick start
Generate TCR-pMHC complex predictions
Generate unbound TCR predictions
Thanks
References
Copyright and license

Quick start

The TCRmodel2 code is adapted from AlphaFold v.2.3.0.

First, clone this repository:

git clone https://github.com/piercelab/tcrmodel2
cd tcrmodel2

Requirements

NVIDIA CUDA driver >= 11.2

Download database

While the majority of database files can be found in data/databases/ folder, due to file size limit, one would need to:

unzip pdb sequence database file:

cd data/databases
tar -xvzf pdb_seqres.txt.tar.gz

download pdb_mmcif and params database (around 120 GB total after unzip) used by alphafold to a database folder of your choice, the path of which will be pass as a ori_db variable to the run_tcrmodel2.py and run_tcrmodel2_ub_tcr.py script. Please refer to the download instructions in download_pdb_mmcif.sh and download_alphafold_params.sh in alphafold repository.

Install Software

To get started with using TCRmodel2, you have two options for installation:

Option 1: Build Singularity Container

This project can be set up using Singularity, which allows you to create and run containers.

Ensure you have Singularity installed on your system. If not, download and install it from the Singularity official website.
Modify elements of the singularity/tcrmodel2_singularity.def like the CUDA version to match your system's configuration.
Build the Singularity container. We offer a pre-built Singularity image file that is compatible with CUDA version 11.2, which you can access here. Please right-click the link and choose 'Save Link As...' to save the file. However, for greater flexibility and compatibility with the CUDA version on your machine, we recommend building the .sif file from the provided .def file. This approach allows you to tailor the build to your specific system requirements.
```
sudo singularity build tcrmodel2.sif singularity/tcrmodel2_singularity.def
```

If you do not have sudo permission, you may build with the following command instead:

singularity build --fakeroot tcrmodel2.sif singularity/tcrmodel2_singularity.def

Run the Singularity container. Example usage can be found in singularity/run_tcrmodel2_singularity.sh. Update variables like ALPHAFOLD_DB, ALPHAFOLD_SIF, OUTPUT_DIR, and elements such as job_id and input sequences tcra_sequence, tcrb_seq, pep_seq, mhca_seq with the appropriate values for your run. For details on how to construct predictions, please refer to sections Generate TCR-pMHC complex predictions and Generate unbound TCR predictions.

Option 2: Step-by-Step Installation

For a manual setup, follow these steps:

Install AlphaFold requirements in a conda environment. Here's a useful resource if you prefer to install AlphaFold without Docker: https://github.com/kalininalab/alphafold_non_docker
Install additional packages: ANARCI and MDAnalysis to the conda environment created from previous step. These two packages are not required for generating structural predictions. ANARCI is used to trim TCR to variable domains only, and for renumbering PDB outputs. MDAnalysis is used for output renumbering and output alignment.
```
conda install -c bioconda anarci
conda config --add channels conda-forge
conda install mdanalysis
```

Generate TCR-pMHC complex predictions

Workflow for creating TCR-pMHC complex structure predictions:

Receive TCR alpha, beta, peptide, MHC sequences
Build pMHC template alignment file
Generate MSA features using a reduced database for all chains, considered seperatedly
Generate all other features by concatenating peptide MHC into one chain
Predict structures
Output 5 structures and a text file containing 1) templates used 2) prediction scores

Peptide length requirement:

For class I TCR-pMHC complexes, kindly ensure that the peptide length is between 8-15.
For class II TCR-pMHC complexes, kindly ensure that the peptide input is 11 aa in length. Specifically, it should consist of a 9 aa core with an additional 1 aa at both the N-terminal and C-terminal of the core peptide.

To make a class I TCR-pMHC prediction:

python run_tcrmodel2.py \
--job_id=test_clsI_6kzw \
--output_dir=experiments/ \
--tcra_seq=AQEVTQIPAALSVPEGENLVLNCSFTDSAIYNLQWFRQDPGKGLTSLLLIQSSQREQTSGRLNASLDKSSGRSTLYIAASQPGDSATYLCAVTNQAGTALIFGKGTTLSVSS \
--tcrb_seq=NAGVTQTPKFQVLKTGQSMTLQCSQDMNHEYMSWYRQDPGMGLRLIHYSVGAGITDQGEVPNGYNVSRSTTEDFPLRLLSAAPSQTSVYFCASSYSIRGSRGEQFFGPGTRLTVL \
--pep_seq=RLPAKAPLL \
--mhca_seq=SHSLKYFHTSVSRPGRGEPRFISVGYVDDTQFVRFDNDAASPRMVPRAPWMEQEGSEYWDRETRSARDTAQIFRVNLRTLRGYYNQSEAGSHTLQWMHGCELGPDGRFLRGYEQFAYDGKDYLTLNEDLRSWTAVDTAAQISEQKSNDASEAEHQRAYLEDTCVEWLHKYLEKGKETLLH \
--ori_db=/path/to/alphafold_database #set it as the path to the folder containing pdb_mmcif and params

To make a class II TCR-pMHC prediction:

python run_tcrmodel2.py \
--job_id=test_clsII_7t2c \
--output_dir=experiments \
--tcra_seq=LAKTTQPISMDSYEGQEVNITCSHNNIATNDYITWYQQFPSQGPRFIIQGYKTKVTNEVASLFIPADRKSSTLSLPRVSLSDTAVYYCLVGDTGFQKLVFGTGTRLLVSP \
--tcrb_seq=GAVVSQHPSWVICKSGTSVKIECRSLDFQATTMFWYRQFPKQSLMLMATSNEGSKATYEQGVEKDKFLINHASLTLSTLTVTSAHPEDSSFYICSARDPGGGGSSYEQYFGPGTRLTVT \
--pep_seq=LAWEWWRTV \
--mhca_seq=IKADHVSTYAAFVQTHRPTGEFMFEFDEDEMFYVDLDKKETVWHLEEFGQAFSFEAQGGLANIAILNNNLNTLIQRSNHTQAT \
--mhcb_seq=PENYLFQGRQECYAFNGTQRFLERYIYNREEFARFDSDVGEFRAVTELGRPAAEYWNSQKDILEEKRAVPDRMCRHNYELGGPMTLQR \
--ori_db=/path/to/alphafold_database #set it as the path to the folder containing pdb_mmcif and params

You may use additional flags in run_tcrmodel2.py to control additional behaviors of the script. To see a list of flags:

python run_tcrmodel2.py --help

Generate unbound TCR predictions

Workflow for creating TCR-pMHC complex structure predictions:

Receive TCR alpha, beta sequences
Generate MSA features using reduced database, and modified TCR template search protocol.
Predict structures
Output 5 structures and a text file containing 1) templates used 2) prediction scores

To make a unbound TCR prediction:

python run_tcrmodel2_ub_tcr.py \
--job_id=test_tcr_7t2b \
--output_dir=experiments \
--tcra_seq=SQQGEEDPQALSIQEGENATMNCSYKTSINNLQWYRQNSGRGLVHLILIRSNEREKHSGRLRVTLDTSKKSSSLLITASRAADTASYFCATDKKGGATNKLIFGTGTLLAVQP \
--tcrb_seq=NAGVTQTPKFRVLKTGQSMTLLCAQDMNHEYMYWYRQDPGMGLRLIHYSVGEGTTAKGEVPDGYNVSRLKKQNFLLGLESAAPSQTSVYFCASSQGGGEQYFGPGTRLTVT \
--ori_db=/path/to/alphafold_database #set it as the path to the folder containing pdb_mmcif and params

You may use additional flags in run_tcrmodel2_ub_tcr.py to control additional behaviors of the script. To see a list of flags:

python run_tcrmodel2_ub_tcr.py --help

Thanks

We would like to thank alphafold, alphafold_finetune, ColabFold teams for developing and distributing the code. The content inside alphafold/ folder is modified from alphafold/ of alphafold repository. The featurization of custom template is modified from predict_utils.py of alphafold_finetune. Chain break introduction, as well as making mock template feature steps are modified from batch.py of ColabFold.

Reference

Yin R, Ribeiro-Filho HV, Lin V, Gowthaman R, Cheung M, Pierce BG. (2023) TCRmodel2: high-resolution modeling of T cell receptor recognition using deep learning. Nucleic Acids Res, 51(W1):W569-W576. https://doi.org/10.1093/nar/gkad356

Copyright and license

Apache License 2.0

piercelab / tcrmodel2