AF2BIND: Prediction of ligand-binding sites using AlphaFold2

Predicting ligand-binding sites, particularly in the absence of previously resolved homologous structures, presents a significant challenge in structural biology. Here, we leverage the internal pairwise representation of AlphaFold2 (AF2) to train a model, AF2BIND, to accurately predict small-molecule-binding residues given only a target protein. AF2BIND uses 20 "bait" amino acids to optimally extract the binding signal in the absence of a small-molecule ligand. We find that the AF2 pair representation outperforms other neural-network representations for binding-site prediction. Moreover, unique combinations of the 20 bait amino acids are correlated with chemical properties of the ligand.

The pipeline for large proteins can be found here: https://github.com/arteemg/af2bind_prod/blob/main/af2bind_large_pdb.ipynb

For more details see preprint:

AF2BIND: Predicting ligand-binding sites using the pair representation of AlphaFold2

• Artem Gazizov, Anna Lian, Casper Alexander Goverde, Sergey Ovchinnikov, Nicholas F. Polizzi
• https://doi.org/10.1101/2023.10.15.562410

Experiments were conducted using the latest ColabDesign github commit v1.1.1, with the Alphafold's weights as of 2022-03-02

System Requirements

Hardware requirement

AF2BIND requires a standalone computer with an Nvidia GPU.

OS Requirements

AF2BIND was tested on Red Hat Enterprise Linux; Version 9.6

Python dependencies

colabdesign
numpy
pandas
jax
pickle
py3Dmol

typical install time ≈ 10 minutes.

Installation

first install jax (with GPU support)

pip install "jax[cuda]" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html

second install colabdesign

pip install git+https://github.com/sokrypton/ColabDesign.git@v1.1.1
# download alphafold weights
mkdir params
curl -fsSL https://storage.googleapis.com/alphafold/alphafold_params_2022-12-06.tar | tar x -C params

Code example

We run af2bind on the PDB file 6w70 to get the predictions. To run the code, you should define the af2bind function, which is located in af2bind.ipynb.

target_pdb = "6w70" 
target_chain = "A"

mask_sidechains = True
mask_sequence = False

target_pdb = target_pdb.replace(" ","")
target_chain = target_chain.replace(" ","")
if target_chain == "":
  target_chain = "A"

pdb_filename = get_pdb(target_pdb)

clear_mem()
af_model = mk_afdesign_model(protocol="binder", debug=True)
af_model.prep_inputs(pdb_filename=pdb_filename,
                     chain=target_chain,
                     binder_len=20,
                     rm_target_sc=mask_sidechains,
                     rm_target_seq=mask_sequence)

r_idx = af_model._inputs["residue_index"][-20] + (1 + np.arange(20)) * 50
af_model._inputs["residue_index"][-20:] = r_idx.flatten()

af_model.set_seq("ACDEFGHIKLMNPQRSTVWY")
af_model.predict(verbose=False)

o = af2bind(af_model.aux["debug"]["outputs"],
            mask_sidechains=mask_sidechains)
pred_bind = o["p_bind"].copy()

License

This project is covered under the MIT License.

Other Models

The *.zip files contain all the other weights for models described in the manuscript (including ESM-2, ESM-IF, combinations, etc.). See train_script.ipynb for details on how features are combined.