Awesome

LigandMPNN

This package provides inference code for LigandMPNN & ProteinMPNN models. The code and model parameters are available under the MIT license.

Third party code: side chain packing uses helper functions from Openfold.

Running the code

git clone https://github.com/dauparas/LigandMPNN.git
cd LigandMPNN
bash get_model_params.sh "./model_params"

#setup your conda/or other environment
#conda create -n ligandmpnn_env python=3.11
#pip3 install -r requirements.txt

python run.py \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/default"

Dependencies

To run the model you will need to have Python>=3.0, PyTorch, Numpy installed, and to read/write PDB files you will need Prody.

For example to make a new conda environment for LigandMPNN run:

conda create -n ligandmpnn_env python=3.11
pip3 install -r requirements.txt

Main differences compared with ProteinMPNN code

• Input PDBs are parsed using Prody preserving protein residue indices, chain letters, and insertion codes. If there are missing residues in the input structure the output fasta file won't have added X to fill the gaps. The script outputs .fasta and .pdb files. It's recommended to use .pdb files since they will hold information about chain letters and residue indices.
• Adding bias, fixing residues, and selecting residues to be redesigned now can be done using residue indices directly, e.g. A23 (means chain A residue with index 23), B42D (chain B, residue 42, insertion code D).
• Model writes to fasta files: overall_confidence, ligand_confidence which reflect the average confidence/probability (with T=1.0) over the redesigned residues overall_confidence=exp[-mean_over_residues(log_probs)]. Higher numbers mean the model is more confident about that sequence. min_value=0.0; max_value=1.0. Sequence recovery with respect to the input sequence is calculated only over the redesigned residues.

Model parameters

To download model parameters run:

bash get_model_params.sh "./model_params"

Available models

To run the model of your choice specify --model_type and optionally the model checkpoint path. Available models:

• ProteinMPNN

--model_type "protein_mpnn"
--checkpoint_protein_mpnn "./model_params/proteinmpnn_v_48_002.pt" #noised with 0.02A Gaussian noise
--checkpoint_protein_mpnn "./model_params/proteinmpnn_v_48_010.pt" #noised with 0.10A Gaussian noise
--checkpoint_protein_mpnn "./model_params/proteinmpnn_v_48_020.pt" #noised with 0.20A Gaussian noise
--checkpoint_protein_mpnn "./model_params/proteinmpnn_v_48_030.pt" #noised with 0.30A Gaussian noise

• LigandMPNN

--model_type "ligand_mpnn"
--checkpoint_ligand_mpnn "./model_params/ligandmpnn_v_32_005_25.pt" #noised with 0.05A Gaussian noise
--checkpoint_ligand_mpnn "./model_params/ligandmpnn_v_32_010_25.pt" #noised with 0.10A Gaussian noise
--checkpoint_ligand_mpnn "./model_params/ligandmpnn_v_32_020_25.pt" #noised with 0.20A Gaussian noise
--checkpoint_ligand_mpnn "./model_params/ligandmpnn_v_32_030_25.pt" #noised with 0.30A Gaussian noise

• SolubleMPNN

--model_type "soluble_mpnn"
--checkpoint_soluble_mpnn "./model_params/solublempnn_v_48_002.pt" #noised with 0.02A Gaussian noise
--checkpoint_soluble_mpnn "./model_params/solublempnn_v_48_010.pt" #noised with 0.10A Gaussian noise
--checkpoint_soluble_mpnn "./model_params/solublempnn_v_48_020.pt" #noised with 0.20A Gaussian noise
--checkpoint_soluble_mpnn "./model_params/solublempnn_v_48_030.pt" #noised with 0.30A Gaussian noise

• ProteinMPNN with global membrane label

--model_type "global_label_membrane_mpnn"
--checkpoint_global_label_membrane_mpnn "./model_params/global_label_membrane_mpnn_v_48_020.pt" #noised with 0.20A Gaussian noise

• ProteinMPNN with per residue membrane label

--model_type "per_residue_label_membrane_mpnn"
--checkpoint_per_residue_label_membrane_mpnn "./model_params/per_residue_label_membrane_mpnn_v_48_020.pt" #noised with 0.20A Gaussian noise

• Side chain packing model

--checkpoint_path_sc "./model_params/ligandmpnn_sc_v_32_002_16.pt"

Design examples

1 default

Default settings will run ProteinMPNN.

python run.py \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/default"

2 --temperature

--temperature 0.05 Change sampling temperature (higher temperature gives more sequence diversity).

python run.py \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --temperature 0.05 \
        --out_folder "./outputs/temperature"

3 --seed

--seed Not selecting a seed will run with a random seed. Running this multiple times will give different results.

python run.py \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/random_seed"

4 --verbose

--verbose 0 Do not print any statements.

python run.py \
        --seed 111 \
        --verbose 0 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/verbose"

5 --save_stats

--save_stats 1 Save sequence design statistics.

#['generated_sequences', 'sampling_probs', 'log_probs', 'decoding_order', 'native_sequence', 'mask', 'chain_mask', 'seed', 'temperature']
python run.py \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/save_stats" \
        --save_stats 1

6 --fixed_residues

--fixed_residues Fixing specific amino acids. This example fixes the first 10 residues in chain C and adds global bias towards A (alanine). The output should have all alanines except the first 10 residues should be the same as in the input sequence since those are fixed.

python run.py \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/fix_residues" \
        --fixed_residues "C1 C2 C3 C4 C5 C6 C7 C8 C9 C10" \
        --bias_AA "A:10.0"

7 --redesigned_residues

--redesigned_residues Specifying which residues need to be designed. This example redesigns the first 10 residues while fixing everything else.

python run.py \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/redesign_residues" \
        --redesigned_residues "C1 C2 C3 C4 C5 C6 C7 C8 C9 C10" \
        --bias_AA "A:10.0"

8 --number_of_batches

Design 15 sequences; with batch size 3 (can be 1 when using CPUs) and the number of batches 5.

python run.py \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/batch_size" \
        --batch_size 3 \
        --number_of_batches 5

9 --bias_AA

Global amino acid bias. In this example, output sequences are biased towards W, P, C and away from A.

python run.py \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --bias_AA "W:3.0,P:3.0,C:3.0,A:-3.0" \
        --out_folder "./outputs/global_bias"

10 --bias_AA_per_residue

Specify per residue amino acid bias, e.g. make residues C1, C3, C5, and C7 to be prolines.

# {
# "C1": {"G": -0.3, "C": -2.0, "P": 10.8},
# "C3": {"P": 10.0},
# "C5": {"G": -1.3, "P": 10.0},
# "C7": {"G": -1.3, "P": 10.0}
# }
python run.py \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --bias_AA_per_residue "./inputs/bias_AA_per_residue.json" \
        --out_folder "./outputs/per_residue_bias"

11 --omit_AA

Global amino acid restrictions. This is equivalent to using --bias_AA and setting bias to be a large negative number. The output should be just made of E, K, A.

python run.py \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --omit_AA "CDFGHILMNPQRSTVWY" \
        --out_folder "./outputs/global_omit"

12 --omit_AA_per_residue

Per residue amino acid restrictions.

# {
# "C1": "ACDEFGHIKLMNPQRSTVW",
# "C3": "ACDEFGHIKLMNPQRSTVW",
# "C5": "ACDEFGHIKLMNPQRSTVW",
# "C7": "ACDEFGHIKLMNPQRSTVW"
# }
python run.py \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --omit_AA_per_residue "./inputs/omit_AA_per_residue.json" \
        --out_folder "./outputs/per_residue_omit"

13 --symmetry_residues

13 --symmetry_weights

Designing sequences with symmetry, e.g. homooligomer/2-state proteins, etc. In this example make C1=C2=C3, also C4=C5, and C6=C7.

#total_logits += symmetry_weights[t]*logits
#probs = torch.nn.functional.softmax((total_logits+bias_t) / temperature, dim=-1)
#total_logits_123 = 0.33*logits_1+0.33*logits_2+0.33*logits_3
#output should be ***ooxx
python run.py \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/symmetry" \
        --symmetry_residues "C1,C2,C3|C4,C5|C6,C7" \
        --symmetry_weights "0.33,0.33,0.33|0.5,0.5|0.5,0.5"

14 --homo_oligomer

Design homooligomer sequences. This automatically sets --symmetry_residues and --symmetry_weights assuming equal weighting from all chains.

python run.py \
        --model_type "ligand_mpnn" \
        --seed 111 \
        --pdb_path "./inputs/4GYT.pdb" \
        --out_folder "./outputs/homooligomer" \
        --homo_oligomer 1 \
        --number_of_batches 2

15 --file_ending

Outputs will have a specified ending; e.g. 1BC8_xyz.fa instead of 1BC8.fa

python run.py \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/file_ending" \
        --file_ending "_xyz"

16 --zero_indexed

Zero indexed names in /backbones/1BC8_0.pdb, 1BC8_1.pdb, 1BC8_2.pdb etc

python run.py \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/zero_indexed" \
        --zero_indexed 1 \
        --number_of_batches 2

17 --chains_to_design

Specify which chains (e.g. "A,B,C") need to be redesigned, other chains will be kept fixed. Outputs in seqs/backbones will still have atoms/sequences for the whole input PDB.

python run.py \
        --model_type "ligand_mpnn" \
        --seed 111 \
        --pdb_path "./inputs/4GYT.pdb" \
        --out_folder "./outputs/chains_to_design" \
        --chains_to_design "A,B"

18 --parse_these_chains_only

Parse and design only specified chains (e.g. "A,B,C"). Outputs will have only specified chains.

python run.py \
        --model_type "ligand_mpnn" \
        --seed 111 \
        --pdb_path "./inputs/4GYT.pdb" \
        --out_folder "./outputs/parse_these_chains_only" \
        --parse_these_chains_only "A,B"

19 --model_type "ligand_mpnn"

Run LigandMPNN with default settings.

python run.py \
        --model_type "ligand_mpnn" \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/ligandmpnn_default"

20 --checkpoint_ligand_mpnn

Run LigandMPNN using 0.05A model by specifying --checkpoint_ligand_mpnn flag.

python run.py \
        --checkpoint_ligand_mpnn "./model_params/ligandmpnn_v_32_005_25.pt" \
        --model_type "ligand_mpnn" \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/ligandmpnn_v_32_005_25"

21 --ligand_mpnn_use_atom_context

Setting --ligand_mpnn_use_atom_context 0 will mask all ligand atoms. This can be used to assess how much ligand atoms affect AA probabilities.

python run.py \
        --model_type "ligand_mpnn" \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/ligandmpnn_no_context" \
        --ligand_mpnn_use_atom_context 0

22 --ligand_mpnn_use_side_chain_context

Use fixed residue side chain atoms as extra ligand atoms.

python run.py \
        --model_type "ligand_mpnn" \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/ligandmpnn_use_side_chain_atoms" \
        --ligand_mpnn_use_side_chain_context 1 \
        --fixed_residues "C1 C2 C3 C4 C5 C6 C7 C8 C9 C10"

23 --model_type "soluble_mpnn"

Run SolubleMPNN (ProteinMPNN-like model with only soluble proteins in the training dataset).

python run.py \
        --model_type "soluble_mpnn" \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/soluble_mpnn_default"

24 --model_type "global_label_membrane_mpnn"

Run global label membrane MPNN (trained with extra input - binary label soluble vs not) --global_transmembrane_label #1 - membrane, 0 - soluble.

python run.py \
        --model_type "global_label_membrane_mpnn" \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/global_label_membrane_mpnn_0" \
        --global_transmembrane_label 0

25 --model_type "per_residue_label_membrane_mpnn"

Run per residue label membrane MPNN (trained with extra input per residue specifying buried (hydrophobic), interface (polar), or other type residues; 3 classes).

python run.py \
        --model_type "per_residue_label_membrane_mpnn" \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/per_residue_label_membrane_mpnn_default" \
        --transmembrane_buried "C1 C2 C3 C11" \
        --transmembrane_interface "C4 C5 C6 C22"

26 --fasta_seq_separation

Choose a symbol to put between different chains in fasta output format. It's recommended to PDB output format to deal with residue jumps and multiple chain parsing.

python run.py \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/fasta_seq_separation" \
        --fasta_seq_separation ":"

27 --pdb_path_multi

Specify multiple PDB input paths. This is more efficient since the model needs to be loaded from the checkpoint once.

#{
#"./inputs/1BC8.pdb": "",
#"./inputs/4GYT.pdb": ""
#}
python run.py \
        --pdb_path_multi "./inputs/pdb_ids.json" \
        --out_folder "./outputs/pdb_path_multi" \
        --seed 111

28 --fixed_residues_multi

Specify fixed residues when using --pdb_path_multi flag.

#{
#"./inputs/1BC8.pdb": "C1 C2 C3 C4 C5 C10 C22",
#"./inputs/4GYT.pdb": "A7 A8 A9 A10 A11 A12 A13 B38"
#}
python run.py \
        --pdb_path_multi "./inputs/pdb_ids.json" \
        --fixed_residues_multi "./inputs/fix_residues_multi.json" \
        --out_folder "./outputs/fixed_residues_multi" \
        --seed 111

29 --redesigned_residues_multi

Specify which residues need to be redesigned when using --pdb_path_multi flag.

#{
#"./inputs/1BC8.pdb": "C1 C2 C3 C4 C5 C10",
#"./inputs/4GYT.pdb": "A7 A8 A9 A10 A12 A13 B38"
#}
python run.py \
        --pdb_path_multi "./inputs/pdb_ids.json" \
        --redesigned_residues_multi "./inputs/redesigned_residues_multi.json" \
        --out_folder "./outputs/redesigned_residues_multi" \
        --seed 111

30 --omit_AA_per_residue_multi

Specify which residues need to be omitted when using --pdb_path_multi flag.

#{
#"./inputs/1BC8.pdb": {"C1":"ACDEFGHILMNPQRSTVWY", "C2":"ACDEFGHILMNPQRSTVWY", "C3":"ACDEFGHILMNPQRSTVWY"},
#"./inputs/4GYT.pdb": {"A7":"ACDEFGHILMNPQRSTVWY", "A8":"ACDEFGHILMNPQRSTVWY"}
#}
python run.py \
        --pdb_path_multi "./inputs/pdb_ids.json" \
        --omit_AA_per_residue_multi "./inputs/omit_AA_per_residue_multi.json" \
        --out_folder "./outputs/omit_AA_per_residue_multi" \
        --seed 111

31 --bias_AA_per_residue_multi

Specify amino acid biases per residue when using --pdb_path_multi flag.

#{
#"./inputs/1BC8.pdb": {"C1":{"A":3.0, "P":-2.0}, "C2":{"W":10.0, "G":-0.43}},
#"./inputs/4GYT.pdb": {"A7":{"Y":5.0, "S":-2.0}, "A8":{"M":3.9, "G":-0.43}}
#}
python run.py \
        --pdb_path_multi "./inputs/pdb_ids.json" \
        --bias_AA_per_residue_multi "./inputs/bias_AA_per_residue_multi.json" \
        --out_folder "./outputs/bias_AA_per_residue_multi" \
        --seed 111

32 --ligand_mpnn_cutoff_for_score

This sets the cutoff distance in angstroms to select residues that are considered to be close to ligand atoms. This flag only affects the num_ligand_res and ligand_confidence in the output fasta files.

python run.py \
        --model_type "ligand_mpnn" \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --ligand_mpnn_cutoff_for_score "6.0" \
        --out_folder "./outputs/ligand_mpnn_cutoff_for_score"

33 specifying residues with insertion codes

You can specify residue using chain_id + residue_number + insersion_code; e.g. redesign only residue B82, B82A, B82B, B82C.

python run.py \
        --seed 111 \
        --pdb_path "./inputs/2GFB.pdb" \
        --out_folder "./outputs/insertion_code" \
        --redesigned_residues "B82 B82A B82B B82C" \
        --parse_these_chains_only "B"

34 parse atoms with zero occupancy

Parse atoms in the PDB files with zero occupancy too.

python run.py \
        --model_type "ligand_mpnn" \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/parse_atoms_with_zero_occupancy" \
        --parse_atoms_with_zero_occupancy 1

Scoring examples

Output dictionary

out_dict = {}
out_dict["logits"] - raw logits from the model
out_dict["probs"] - softmax(logits)
out_dict["log_probs"] - log_softmax(logits)
out_dict["decoding_order"] - decoding order used (logits will depend on the decoding order)
out_dict["native_sequence"] - parsed input sequence in integers
out_dict["mask"] - mask for missing residues (usually all ones)
out_dict["chain_mask"] - controls which residues are decoded first
out_dict["alphabet"] - amino acid alphabet used
out_dict["residue_names"] - dictionary to map integers to residue_names, e.g. {0: "C10", 1: "C11"}
out_dict["sequence"] - parsed input sequence in alphabet
out_dict["mean_of_probs"] - averaged over batch_size*number_of_batches probabilities, [protein_length, 21]
out_dict["std_of_probs"] - same as above, but std

1 autoregressive with sequence info

Get probabilities/scores for backbone-sequence pairs using autoregressive probabilities: p(AA_1|backbone), p(AA_2|backbone, AA_1) etc. These probabilities will depend on the decoding order, so it's recomended to set number_of_batches to at least 10.

python score.py \
        --model_type "ligand_mpnn" \
        --seed 111 \
        --autoregressive_score 1\
        --pdb_path "./outputs/ligandmpnn_default/backbones/1BC8_1.pdb" \
        --out_folder "./outputs/autoregressive_score_w_seq" \
        --use_sequence 1\
        --batch_size 1 \
        --number_of_batches 10

2 autoregressive with backbone info only

Get probabilities/scores for backbone using probabilities: p(AA_1|backbone), p(AA_2|backbone) etc. These probabilities will depend on the decoding order, so it's recomended to set number_of_batches to at least 10.

python score.py \
        --model_type "ligand_mpnn" \
        --seed 111 \
        --autoregressive_score 1\
        --pdb_path "./outputs/ligandmpnn_default/backbones/1BC8_1.pdb" \
        --out_folder "./outputs/autoregressive_score_wo_seq" \
        --use_sequence 0\
        --batch_size 1 \
        --number_of_batches 10

3 single amino acid score with sequence info

Get probabilities/scores for backbone-sequence pairs using single aa probabilities: p(AA_1|backbone, AA_{all except AA_1}), p(AA_2|backbone, AA_{all except AA_2}) etc. These probabilities will depend on the decoding order, so it's recomended to set number_of_batches to at least 10.

python score.py \
        --model_type "ligand_mpnn" \
        --seed 111 \
        --single_aa_score 1\
        --pdb_path "./outputs/ligandmpnn_default/backbones/1BC8_1.pdb" \
        --out_folder "./outputs/single_aa_score_w_seq" \
        --use_sequence 1\
        --batch_size 1 \
        --number_of_batches 10

4 single amino acid score with backbone info only

Get probabilities/scores for backbone-sequence pairs using single aa probabilities: p(AA_1|backbone), p(AA_2|backbone) etc. These probabilities will depend on the decoding order, so it's recomended to set number_of_batches to at least 10.

python score.py \
        --model_type "ligand_mpnn" \
        --seed 111 \
        --single_aa_score 1\
        --pdb_path "./outputs/ligandmpnn_default/backbones/1BC8_1.pdb" \
        --out_folder "./outputs/single_aa_score_wo_seq" \
        --use_sequence 0\
        --batch_size 1 \
        --number_of_batches 10

Side chain packing examples

1 design a new sequence and pack side chains (return 1 side chain packing sample - fast)

Design a new sequence using any of the available models and also pack side chains of the new sequence. Return only a single solution for the side chain packing.

python run.py \
        --model_type "ligand_mpnn" \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/sc_default_fast" \
        --pack_side_chains 1 \
        --number_of_packs_per_design 0 \
        --pack_with_ligand_context 1

2 design a new sequence and pack side chains (return 4 side chain packing samples)

Same as above, but returns 4 independent samples for side chains. b-factor shows log prob density per chi angle group.

python run.py \
        --model_type "ligand_mpnn" \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/sc_default" \
        --pack_side_chains 1 \
        --number_of_packs_per_design 4 \
        --pack_with_ligand_context 1

3 fix specific residues fors sequence design and packing

This option will not repack side chains of the fixed residues, but use them as a context.

python run.py \
        --model_type "ligand_mpnn" \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/sc_fixed_residues" \
        --pack_side_chains 1 \
        --number_of_packs_per_design 4 \
        --pack_with_ligand_context 1 \
        --fixed_residues "C6 C7 C8 C9 C10 C11 C12 C13 C14 C15" \
        --repack_everything 0

4 fix specific residues for sequence design but repack everything

This option will repacks all the residues.

python run.py \
        --model_type "ligand_mpnn" \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/sc_fixed_residues_full_repack" \
        --pack_side_chains 1 \
        --number_of_packs_per_design 4 \
        --pack_with_ligand_context 1 \
        --fixed_residues "C6 C7 C8 C9 C10 C11 C12 C13 C14 C15" \
        --repack_everything 1

5 design a new sequence using LigandMPNN but pack side chains without considering ligand/DNA etc atoms

You can run side chain packing without taking into account context atoms like DNA atoms. This most likely will results in side chain clashing with context atoms, but it might be interesting to see how model's uncertainty changes when ligand atoms are present vs not for side chain conformations.

python run.py \
        --model_type "ligand_mpnn" \
        --seed 111 \
        --pdb_path "./inputs/1BC8.pdb" \
        --out_folder "./outputs/sc_no_context" \
        --pack_side_chains 1 \
        --number_of_packs_per_design 4 \
        --pack_with_ligand_context 0

Things to add

• Support for ProteinMPNN CA-only model.
• Examples for scoring sequences only.
• Side-chain packing scripts.
• TER

Citing this work

If you use the code, please cite:

@article{dauparas2023atomic,
  title={Atomic context-conditioned protein sequence design using LigandMPNN},
  author={Dauparas, Justas and Lee, Gyu Rie and Pecoraro, Robert and An, Linna and Anishchenko, Ivan and Glasscock, Cameron and Baker, David},
  journal={Biorxiv},
  pages={2023--12},
  year={2023},
  publisher={Cold Spring Harbor Laboratory}
}

@article{dauparas2022robust,
  title={Robust deep learning--based protein sequence design using ProteinMPNN},
  author={Dauparas, Justas and Anishchenko, Ivan and Bennett, Nathaniel and Bai, Hua and Ragotte, Robert J and Milles, Lukas F and Wicky, Basile IM and Courbet, Alexis and de Haas, Rob J and Bethel, Neville and others},
  journal={Science},
  volume={378},
  number={6615},  
  pages={49--56},
  year={2022},
  publisher={American Association for the Advancement of Science}
}