This repository is dedicated to the BioML Challenge 2024: Bits to Binders competition, organized by the University of Texas at Austin BioML Society. The aim of this competition is to advance protein design using modern AI tools. Our task is to design a protein sequence that effectively binds to a specific cancer antigen target, with the goal of activating CAR-T cell proliferation and killing responses.
To run the workflow, you need to install RFdiffusion, ColabDesign, and pyrosetta.
Here's a step-by-step breakdown of the workflow used to design and validate the protein binders:
We were constrained to designing binders with a maximum length of 80 amino acids (AA). While smaller binders could be extended with linkers, we aimed to engineer larger binders at the full 80 AA, giving us more flexibility and space for engineering.
To achieve this, we developed a BinderBlueprint generator to create different adjacency matrices for binders of a defined length and secondary structural elements (SSEs). These matrices are later used with RFdiffusion for scaffold-guided diffusion of binders towards desired target hotspots.
Follow the example below to generate binder blueprints, or run the provided script scripts/binder_blueprints.py to generate binder blueprints for different scaffolds - binders with different structural elements.
If you want to provide just the size of the binder (with the length of beta strands being half that of helices):
blueprint = BinderBlueprints(
elements=["HHH"], # Three helices
size=80,
)
blueprint.save_adj_sse(output_dir=OUTPUT_DIR)You can also specify the sizes of the individual secondary elements and linkers:
blueprint = BinderBlueprints(
elements=["HHH"], # Three helices
size=80,
element_lengths=[25,25,24], # Lengths of each helix
linker_lengths=[3,3], # Lengths of the linkers between the helices
)
blueprint.save_adj_sse(output_dir=OUTPUT_DIR)This will generate secondary structure (ss.pt) and adjacency matrices (adj.pt) for the binders, which can be used as input for scaffold-guided diffusion.
To use our pre-generated binder blueprints, extract them with the following command:
tar -xzvf data/binder_blueprints.tar.gzWe employed several approaches to design the protein binders, leveraging various RFdiffusion scripts targeting different regions of the hCD20 antigen:
Targeting a single monomer (modeled as a dimer) Targeting the middle of hCD20 Targeting the entire dimer with random binder scaffolds Preparing symmetric binders to leverage target symmetry
We shorten the hCD20 protein model to only the part that is positioned on the outer site of membrane - contigs: D65-87/D131-198/0 C65-87/C131-198/0. This is so that computations will be quicker. We also need to generate ss.pt and adj.pt files for RFdiffusion, easiest way is to use make_secstruc_adj.py from RFdiffusion repository.
The only difference between the first two approaches is the choice of diffusion hotspots:
Targeting a single hCD20 monomer: [D76,D160,D161,D163,D166,D170,C177,C178,C182]
Targeting the middle region of hCD20: [D77,D161,D163,D166,D171,D174,D178,C77,C161,C163,C166,C171,C174,C178]
python /RFdiffusion_path/run_inference.py \
diffuser.T=50 \
inference.output_prefix={output} \
scaffoldguided.target_path={target} \
scaffoldguided.scaffoldguided=True \
ppi.hotspot_res=[{hotspots}] \
scaffoldguided.target_pdb=True \
scaffoldguided.target_ss={ss} \
scaffoldguided.target_adj={adj} \
scaffoldguided.scaffold_dir={scaffolds} \
potentials.guiding_potentials=[ \
"type:binder_ROG,weight:3", \
"type:interface_ncontacts", \
"type:binder_distance_ReLU" \
] \
potentials.guide_scale=2 \
potentials.guide_decay="quadratic" \
inference.num_designs={num_of_diffusions} \
denoiser.noise_scale_ca=0 \
denoiser.noise_scale_frame=0For targeting the entire hCD20 dimer, we use the same hotspot residues but specify the contigs, as we are not doing scaffold guided diffusion.
Hotspots: [D77,D161,D163,D166,D171,D174,D178,C77,C161,C163,C166,C171,C174,C178]
Contigs: D65-87/D131-198/0 C65-87/C131-198/0 80-80
python /RFdiffusion_path/run_inference.py \
diffuser.T=50 \
inference.output_prefix={output} \
inference.input_pdb={target} \
contigmap.contigs=[{contigs}] \
ppi.hotspot_res=[{hotspots}] \
potentials.guiding_potentials=[ \
"type:binder_ROG,weight:3", \
"type:interface_ncontacts", \
"type:binder_distance_ReLU" \
] \
potentials.guide_scale=2 \
potentials.guide_decay="quadratic" \
inference.num_designs={num_of_diffusions}
To exploit the C2 symmetry of hCD20, we generated symmetric binders. This required a small modification to RFdiffusion to allow for non-receptor hotspots.
The contigs specify the symmetry relations as: A65-87/A131-198 80-80/0 A65-87/A131-198 80-80/0. For symmetric designs, we adjusted the potentials to avoid designs that did not interact properly, often using lower contact weights ("type:olig_contacts,weight_intra:0.1,weight_inter:0.2") or even skipping guiding potentials.
python /RFdiffusion_path/run_inference.py \
--config-name=symmetry \
diffuser.T=20\
inference.input_pdb={target} \
inference.num_designs={num_of_diffusions} \
inference.symmetry="C2" \
inference.ckpt_override_path=/RFdiffusion_path/models/Complex_beta_ckpt.pt \
inference.output_prefix={output} \
contigmap.contigs=[{contigs}] \
ppi.hotspot_res=[{hotspots}] \
potentials.guiding_potentials=[{guiding_potentials}] \
potentials.olig_intra_all=True \
potentials.olig_inter_all=True \
potentials.guide_scale=2 \
potentials.guide_decay="quadratic"To validate the generated protein binders, we employed a two-step process combining ProteinMPNN for sequence design and AlphaFold2 for structural validation.
For non-symmetric binders, use the provided script scripts/mpnn_af2.py, which integrates ProteinMPNN for sequence generation and AlphaFold2 for structural validation:
python scripts/mpnn_af2.py {pdb} {output} {target_chain} {binder_chain} \
--sampling_temp {sampling_temp} \
--num_recycles {num_recycles} \
--num_seqs {num_seqs} \
--num_filt_seq {num_filtered_seqs} \
--results_dataframe {output_folder} \
--save_best_onlyFor symmetric binders, use the script scripts/mpnn_af2_sym.py, which is adapted for symmetric designs:
python scripts/mpnn_af2_sym.py {pdb} {output} {contig} \
--num_seqs {num_seqs} \
--sampling_temp {sampling_temp} \
--num_recycles {num_recycles} \
--rm_aa="C" \
--copies=2 \
--mpnn_batch {mpnn_batch} \
--results_dataframe {output_dir}/sym_diff \
--save_best_only --initial_guess \
--use_multimer --use_solubleTo explore variations in binder design, we used the scripts/binder_redesign.py script, which utilizes the ColabDesign library. This script allows for the refinement of protein binders or the generation of binder scaffold variants that can be used as input for partial diffusion.
python /scripts/binder_redesign.py \
input_pdb={pdb} \
--num_recycles={number_of_recycles} \
--num_models={number_of_models} \
--recycle_mode={recycle_mode_choice} \
--target_chain={target_chain} \
--target_flexible \
--binder_chain={binder_chain} \
--soft_iters={soft_iterations} \
--hard_iters={hard_iterations} \
--output_file={output_pdb} \
--data_dir={model_data_directory} # path to alphafold2 paramsTo further explore the structural space and related binder scaffolds, we applied partial diffusion to the best designed binders. This allowed us generate variations of the original designs.
python /RFdiffusion_path/run_inference.py \
inference.input_pdb={pdb} \
contigmap.contigs=[{contigs}] \
inference.num_designs={number_of_diffusions} \
diffuser.partial_T={steps}These newly refined designs were subsequently validated using ProteinMPNN and AlphaFold2.
We can also apply partial diffusion while taking advantage of C2 symmetry. To do this, the input needs to be an asymmetrical unit PDB file, with the binder in chain A and the target in chain B (processed from an already predicted complex).
For example, the contig can be specified as: 80-80/0 B1-23/B67-134 80-80/0 B1-23/B67-134.
python /RFdiffusion_path/run_inference.py \
--config-name=symmetry \
inference.symmetry="C2" \
inference.output_prefix={output_prefix}
inference.input_pdb={pdb} \
contigmap.contigs=[{contigs}] \
inference.num_designs={number_of_diffusions} \
diffuser.partial_T={steps} \
potentials.olig_intra_all=True \
potentials.olig_inter_all=True \
potentials.guide_scale=2 \
potentials.guide_decay="quadratic"These newly refined designs can then be validated using ProteinMPNN and AlphaFold2.
For the final selection of binders, we used PyRosetta to compute a range of structural and energetic metrics, helping us identify the most promising candidates. The script scripts/pyrosetta_metrics.py generates these metrics, with key ones including ddg, rg, charge, sap, shape_comp, cms, ddg_cms, and vbuns.
python scripts/pyrosetta_metrics.py {pdb} {target_chain} {binder_chain} {output_dataframe} {xml_file}NOTE: The output dataframe from AlphaFold2 validation can be reused, with new values appended for each protein.
This project is open-source and available under the MIT License. Feel free to explore, modify, and build upon it. For more details, see the LICENSE file.
For questions, suggestions, or collaborations, feel free to reach out [email protected]
@article{Watson2023,
title = {De novo design of protein structure and function with RFdiffusion},
author = {Watson, Joseph L. and Juergens, David and Bennett, Nathaniel R. and Trippe, Brian L. and Yim, Jason and Eisenach, Helen E. and Ahern, Woody and Borst, Andrew J. and Ragotte, Robert J. and Milles, Lukas F. and Wicky, Basile I. M. and Hanikel, Nikita and Pellock, Samuel J. and Courbet, Alexis and Sheffler, William and Wang, Jue and Venkatesh, Preetham and Sappington, Isaac and Torres, Susana Vázquez and Lauko, Anna and De Bortoli, Valentin and Mathieu, Emile and Ovchinnikov, Sergey and Barzilay, Regina and Jaakkola, Tommi S. and DiMaio, Frank and Baek, Minkyung and Baker, David},
DOI = {10.1038/s41586-023-06415-8},
journal = {Nature},
year = {2023},
month = jul
}@article{Dauparas2022,
title = {Robust deep learning–based protein sequence design using ProteinMPNN},
author = {Dauparas, J. and Anishchenko, I. and Bennett, N. and Bai, H. and Ragotte, R. J. and Milles, L. F. and Wicky, B. I. M. and Courbet, A. and de Haas, R. J. and Bethel, N. and Leung, P. J. Y. and Huddy, T. F. and Pellock, S. and Tischer, D. and Chan, F. and Koepnick, B. and Nguyen, H. and Kang, A. and Sankaran, B. and Bera, A. K. and King, N. P. and Baker, D.},
DOI = {10.1126/science.add2187},
journal = {Science},
year = {2022},
month = oct
}@article{Evans2021,
title = {Protein complex prediction with AlphaFold-Multimer},
author = {Evans, Richard and O’Neill, Michael and Pritzel, Alexander and Antropova, Natasha and Senior, Andrew and Green, Tim and Žídek, Augustin and Bates, Russ and Blackwell, Sam and Yim, Jason and Ronneberger, Olaf and Bodenstein, Sebastian and Zielinski, Michal and Bridgland, Alex and Potapenko, Anna and Cowie, Andrew and Tunyasuvunakool, Kathryn and Jain, Rishub and Clancy, Ellen and Kohli, Pushmeet and Jumper, John and Hassabis, Demis},
DOI = {10.1101/2021.10.04.463034},
journal = {bioRxiv},
year = {2021},
month = oct
}@article{Mirdita2022,
title = {ColabFold: making protein folding accessible to all},
author = {Mirdita, Milot and Sch\"{u}tze, Konstantin and Moriwaki, Yoshitaka and Heo, Lim and Ovchinnikov, Sergey and Steinegger, Martin},
DOI = {10.1038/s41592-022-01488-1},
journal = {Nature Methods},
year = {2022},
month = may
}@article{Jumper2021,
title = {Highly accurate protein structure prediction with AlphaFold},
author = {Jumper, John and Evans, Richard and Pritzel, Alexander and Green, Tim and Figurnov, Michael and Ronneberger, Olaf and Tunyasuvunakool, Kathryn and Bates, Russ and Žídek, Augustin and Potapenko, Anna and Bridgland, Alex and Meyer, Clemens and Kohl, Simon A. A. and Ballard, Andrew J. and Cowie, Andrew and Romera-Paredes, Bernardino and Nikolov, Stanislav and Jain, Rishub and Adler, Jonas and Back, Trevor and Petersen, Stig and Reiman, David and Clancy, Ellen and Zielinski, Michal and Steinegger, Martin and Pacholska, Michalina and Berghammer, Tamas and Bodenstein, Sebastian and Silver, David and Vinyals, Oriol and Senior, Andrew W. and Kavukcuoglu, Koray and Kohli, Pushmeet and Hassabis, Demis},
DOI = {10.1038/s41586-021-03819-2},
journal = {Nature},
year = {2021},
month = jul
}@article{Chaudhury2010,
title = {PyRosetta: a script-based interface for implementing molecular modeling algorithms using Rosetta},
author = {Chaudhury, Sidhartha and Lyskov, Sergey and Gray, Jeffrey J.},
DOI = {10.1093/bioinformatics/btq007},
journal = {Bioinformatics},
year = {2010},
month = jan
}