ImmuneFold

Introduction

ImmuneFold is an advanced approach for predicting immune protein structures using transfer learning, adapted from state-of-the-art protein structure prediction frameworks. ImmuneFold is specifically tailored for accurate modeling of immune proteins, including T-cell receptors (TCRs), antibodies, nanobodies, and their complexes with target antigens. By providing precise predictions of immune protein-antigen pairings, ImmuneFold offers valuable insights into protein interaction mechanisms, thereby supporting applications such as vaccine development and immune response analysis.

Installation

To install ImmuneFold, the recommended method is to create a conda environment and install the required dependencies:

git clone git@github.com:CarbonMatrixLab/immunefold.git 
conda env create -f environment.yml
pip install fair-esm

ImmuneFold also need to install the ANARCI. Additionally, to compute the mutation effects on TCR-pMHC interactions, please install the PyRosetta package.

Model Weights

1. Download the ImmuneFold-TCR and ImmuneFold-Ab here and place them in the ./params directory.
2. Download the ESM2 model weights from this link and the contact regressor weights from here. Save these files in the ./params directory.

Usage

TCR-pMHC Structure Prediction

To predict the structure of TCR-pMHC complexes, provide the TCR, peptide, and MHC sequences in a FASTA file, TCR_B_A_P_M.fasta, where B, A, P and M denote the beta, alpha, peptide, MHC chain ids, respectively. The format is as follows Beta_seq:Alpha_seq:Peptide_seq:MHC_seq:

>TCR_B_A_P_M
VSQHPSWVICKSGTSVKIECRSLDFQATTMFWYRQFPKQSLMLMATSNEGSKATYEQGVEKDKFLINHASLTLSTLTVTSAHPEDSSFYICSVSRDRNTGELFFGEGSRLTVL:VEQDPGPFNVPEGATVAFNCTYSNSASQSFFWYRQDCRKEPKLLMSVYSSGNEDGRFTAQLNRASQYISLLIRDSKLSDSATYLCVVNEEDALIFGKGTTLSVSS:YLQPRTFLL:GSHSMRYFFTSVSRPGRGEPRFIAVGYVDDTQFVRFDSDAASQRMEPRAPWIEQEGPEYWDGETRKVKAHSQTHRVDLGTLRGYYNQSEAGSHTVQRMYGCDVGSDWRFLRGYHQYAYDGKDYIALKEDLRSWTAADMAAQTTKHKWEAAHVAEQLRAYLEGTCVEWLRRYLENGKETLQ

To run the ImmuneFold-TCR structure prediction model, use the following command:

python inference.py --config-name=TCR_structure_prediction

Zero-Shot Binding Affinity Prediction for TCR-pMHC

To predict binding affinity using zero-shot learning, provide the whole TCR-pMHC sequences in a FASTA file, following the same format as described for structure prediction:

python inference.py --config-name=TCR_structure_prediction
python predict_energy.py --pdb_dir /path/to/pdb/dir --name_idx /path/to/name_idx.idx --output_file /path/to/energy.csv --mode interface

Unbound Antibody or Nanobody Structure Prediction

For antibody or nanobody structure prediction, provide the sequences in FASTA files, antibody_H_L.fasta or nanobody_H.fasta, where H and L represent the heavy and light chain ids, respectively. The formats are as follows:

Antibody:

>antibody_H_L
VSQHPSWVICKSGTSVKIECRSLDFQATTMFWYRQFPKQSLMLMATSNEGSKATYEQGVEKDKFLINHASLTLSTLTVTSAHPEDSSFYICSVSRDRNTGELFFGEGSRLTVL:VEQDPGPFNVPEGATVAFNCTYSNSASQSFFWYRQDCRKEPKLLMSVYSSGNEDGRFTAQLNRASQYISLLIRDSKLSDSATYLCVVNEEDALIFGKGTTLSVSS

Nanobody:

>nanobody_H
VSQHPSWVICKSGTSVKIECRSLDFQATTMFWYRQFPKQSLMLMATSNEGSKATYEQGVEKDKFLINHASLTLSTLTVTSAHPEDSSFYICSVSRDRNTGELFFGEGSRLTVL

To run the model:

python inference.py --config-name=antibody_structure_prediction

or for nanobodies:

python inference.py --config-name=nanobody_structure_prediction

Bound Antibody or Nanobody Structure Prediction with Target Antigen

For predicting antibody or nanobody structures bound to a target antigen, provide the antigen structure as a PDB file antigen.pdb along with the antibody or nanobody sequences, and run the following command:

python inference.py --config-name=antibody_antigen_structure_prediction

Synthetic TCR Dataset in VDJdb

To showcase the capabilities of ImmuneFold in large-scale TCR structure prediction, we applied the model to the VDJdb dataset, comprising 32,703 non-redundant TCR sequences. The predicted structures are available for download at link.