BayesDesign

BayesDesign is an algorithm for designing proteins with high stability and conformational specificity. See preprint here.

Try out the BayesDesign model here:

Dependencies: ./dependencies/requirements.txt.

One-line sequence design

To design a protein sequence to fit a protein backbone:

python3 design.py --model_name bayes_design --protein_id 6MRR --decode_order n_to_c --decode_algorithm beam --n_beams 128 --fixed_positions 67 68

Detailed steps to run with Docker

Clone repository

git clone https://github.com/dellacortelab/bayes_design.git

Build docker image (should take ~5 minutes)

docker build -t bayes_design -f ./bayes_design/dependencies/Dockerfile ./bayes_design/dependencies

Run container

docker run -dit --gpus all --name bayes_dev --rm -v $(pwd)/bayes_design:/code -v $(pwd)/bayes_design/data:/data bayes_design
docker exec -it bayes_dev /bin/bash

Redesign a protein backbone

cd ./code && python3 design.py --model_name bayes_design --protein_id 6MRR --decode_order n_to_c --decode_algorithm beam --n_beams 128 --fixed_positions 67 68

On a V100 GPU, the greedy algorithm predicts ~10 residues/s and beam search with 128 beams predicts 1 residue every 2s.

Citation

@Article{Stern2023,
author={Stern, Jacob A. and Free, Tyler J. and Stern, Kimberlee L. and Gardiner, Spencer and Dalley, Nicholas A. and Bundy, Bradley C. and Price, Joshua L. and Wingate, David and Della Corte, Dennis},
title={A probabilistic view of protein stability, conformational specificity, and design},
journal={Scientific Reports},
year={2023},
volume={13},
number={1},
pages={15493},
issn={2045-2322},
doi={10.1038/s41598-023-42032-1},
url={https://doi.org/10.1038/s41598-023-42032-1}
}

Experiments on designed sequences

Evaluate the probability of a designed sequence under a probability model

python3 experiment.py compare_seq_metric --metric log_prob --protein_id 1PIN --model_name bayes_design --decode_order n_to_c --fixed_positions 34 34 --sequences MLPEGWVKQRNPITGEDVCFNTLTHEMTKFEPQG

python3 experiment.py make_pssm --protein_id 1PIN --model_name pssm --pssm_path ./results/bayes_design_1PIN_pssm.pkl --decode_order n_to_c --fixed_positions 34 34 --sequences ./results/bayes_design_1PIN_sequences.txt

python3 experiment.py make_hist --protein_id 1PIN --model_name pssm --pssm_path ./results/bayes_design_1PIN_pssm.pkl --metric log_prob --decode_order n_to_c --fixed_positions 34 34 --sequences_path ./results/bayes_design_1PIN_sequences.txt

python3 experiment.py seq_filter --protein_id 1PIN --model_name pssm --pssm_path ./results/bayes_design_1PIN_pssm.pkl --metric log_prob --decode_order n_to_c --fixed_positions 34 34 --sequences_path ./results/bayes_design_1PIN_sequences.txt

BayesDesign WW python3 experiment.py seq_filter --protein_id 1PIN --model_name pssm --pssm_path ./results/bayes_design_1PIN_pssm.pkl --metric log_prob --decode_order n_to_c --fixed_positions 34 34 --sequences_path ./results/bayes_design_1PIN_sequences.txt --n_seqs 3

Result: [(-60.534150819590586, 'MLPQGWQAKQDRDTNQWVYRNWITNKITFNKPRG'), (-62.07557004996536, 'KLPEGWIETKDVIHGKTQYHNVNLNETMEEQPVG'), (-62.343200271562175, 'ALIEVWQKQKDPETGQTKYLNVGKGERTPQRPKG')]

ProteinMPNN WW python3 experiment.py seq_filter --protein_id 1PIN --model_name pssm --pssm_path ./results/protein_mpnn_1PIN_pssm.pkl --metric log_prob --decode_order n_to_c --fixed_positions 34 34 --sequences_path ./results/protein_mpnn_1PIN_sequences.txt --n_seqs 3

Result: [(-43.503443719475136, 'ALPTGWEEKIDPVTNQLIYYNVKTKETTTEKPVG'), (-43.567347443909554, 'ELPEGWVEMVDIKTGEVVYYNDITKEITKEKPVG'), (-45.31501494556361, 'ALPAGWEEIIDPETGKVQYYNSQTKEVTTARPIG')]

ProteinMPNN NanoLuc Full Redesign python3 design.py --model_name protein_mpnn --protein_id nanoluc --decode_order n_to_c --decode_algorithm sample --temperature 1. --n_designs 1000 --fixed_positions 1 9

python3 experiment.py make_pssm --pssm_path ./results/protein_mpnn_nanoluc_full_pssm.pkl --sequences_path ./results/protein_mpnn_nanoluc_full_sequences.txt

python3 experiment.py seq_filter --protein_id nanoluc --model_name pssm --pssm_path ./results/protein_mpnn_nanoluc_full_pssm.pkl --metric log_prob --decode_order n_to_c --fixed_positions 1 9 --sequences_path ./results/protein_mpnn_nanoluc_full_sequences.txt --n_seqs 10

ProteinMPNN NanoLuc Partial Redesign python3 design.py --model_name protein_mpnn --protein_id nanoluc --decode_order n_to_c --decode_algorithm sample --temperature 1. --n_designs 1000 --fixed_positions 1 62 97 179

python3 experiment.py make_pssm --pssm_path ./results/protein_mpnn_nanoluc_partial_pssm.pkl --sequences_path ./results/protein_mpnn_nanoluc_partial_sequences.txt

python3 experiment.py seq_filter --protein_id nanoluc --model_name pssm --pssm_path ./results/protein_mpnn_nanoluc_partial_pssm.pkl --metric log_prob --decode_order n_to_c --fixed_positions 1 62 97 179 --sequences_path ./results/protein_mpnn_nanoluc_partial_sequences.txt --n_seqs 10

BayesDesign NanoLuc Full Redesign python3 design.py --model_name bayes_design --protein_id nanoluc --decode_order n_to_c --decode_algorithm sample --temperature 1. --n_designs 1000 --fixed_positions 1 9

python3 experiment.py make_pssm --pssm_path ./results/bayes_design_nanoluc_full_pssm.pkl --sequences_path ./results/bayes_design_nanoluc_full_sequences.txt

python3 experiment.py seq_filter --protein_id nanoluc --model_name pssm --pssm_path ./results/bayes_design_nanoluc_full_pssm.pkl --metric log_prob --decode_order n_to_c --fixed_positions 1 9 --sequences_path ./results/bayes_design_nanoluc_full_sequences.txt --n_seqs 10

BayesDesign NanoLuc Partial Redesign python3 design.py --model_name bayes_design --protein_id nanoluc --decode_order n_to_c --decode_algorithm sample --temperature 1. --n_designs 1000 --fixed_positions 1 62 97 179

python3 experiment.py make_pssm --pssm_path ./results/bayes_design_nanoluc_partial_pssm.pkl --sequences_path ./results/bayes_design_nanoluc_partial_sequences.txt

python3 experiment.py seq_filter --protein_id nanoluc --model_name pssm --pssm_path ./results/bayes_design_nanoluc_partial_pssm.pkl --metric log_prob --decode_order n_to_c --fixed_positions 1 62 97 179 --sequences_path ./results/bayes_design_nanoluc_partial_sequences.txt --n_seqs 10

ProteinMPNN NanoLuc Partial Redesign python3 design.py --model_name bayes_design --protein_id nanoluc --decode_order n_to_c --decode_algorithm max_prob_decode --n_designs 1 --fixed_positions 1 47 54 83 93 167 172 179