File size: 5,697 Bytes
722e0b0 9f054b3 722e0b0 9f054b3 722e0b0 9f054b3 722e0b0 9f054b3 722e0b0 9f054b3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 | ---
language:
- en
tags:
- biology
- protein-design
- inverse-folding
- ligandmpnn
- proteinmpnn
- bindmpnn
---
# BindMPNN
BindMPNN provides **LigandMPNN-compatible and ProteinMPNN-compatible checkpoints** for amino-acid sequence design from protein structures. Use them with the [official LigandMPNN implementation](https://github.com/dauparas/LigandMPNN); the model options and examples for each architecture are given below.
## LigandMPNN-compatible checkpoints
| Checkpoint | Backbone noise level | Protein neighbors | Atom context size |
|---|---:|---:|---:|
| `BindMPNN_v_32_010_25.pt` | 0.10 Å | 32 | 25 |
| `BindMPNN_v_32_020_25.pt` | 0.20 Å | 32 | 25 |
The noise level identifies the checkpoint variant; it is **not** the sampling temperature and does not require adding noise to your input structure. Both files contain inference weights and the metadata required by LigandMPNN's checkpoint loader.
## Installation and download
Use a Python 3.11 environment and install LigandMPNN according to its [installation instructions](https://github.com/dauparas/LigandMPNN#dependencies). For example:
```bash
git clone https://github.com/dauparas/LigandMPNN.git
cd LigandMPNN
python -m pip install -r requirements.txt
python -m pip install huggingface_hub
```
Download the checkpoints into this checkout:
```python
from huggingface_hub import hf_hub_download
for filename in ("BindMPNN_v_32_010_25.pt", "BindMPNN_v_32_020_25.pt"):
hf_hub_download(
repo_id="MurrellLab/SeqGivenStruct",
filename=filename,
local_dir="checkpoints",
)
```
## Sequence design
This example redesigns chain **B** of a target–binder complex while keeping the amino-acid identities of the other chains fixed:
```bash
python run.py \
--model_type ligand_mpnn \
--checkpoint_ligand_mpnn checkpoints/BindMPNN_v_32_010_25.pt \
--pdb_path complex.pdb \
--chains_to_design B \
--ligand_mpnn_use_side_chain_context 1 \
--temperature 0.1 \
--seed 123 \
--batch_size 1 \
--number_of_batches 8 \
--out_folder outputs/bindmpnn
```
Replace `B` with the chain you want to redesign and `complex.pdb` with your input structure. To use the 0.20 Å variant, change the checkpoint filename to `BindMPNN_v_32_020_25.pt`.
- `--chains_to_design` specifies the mutable chains; other chains remain fixed.
- `--ligand_mpnn_use_side_chain_context 1` exposes sidechain atoms of fixed residues as context. Use `0` to omit that sidechain context.
- `--temperature` controls sequence sampling diversity independently of the checkpoint's noise label. The example uses 0.1; adjust it for your application.
- `--seed` controls reproducibility. Sequence files are written under the output folder's `seqs/` directory.
These are sequence-design checkpoints. Sidechain packing uses LigandMPNN's separate packing checkpoint and options; consult the upstream documentation for that workflow and for residue-level constraints, scoring, and batching.
## Programmatic loading
The checkpoint files are PyTorch dictionaries with `model_state_dict`, `num_edges`, `atom_context_num`, and `noise_level`. A compatible loader can use:
```python
import torch
checkpoint = torch.load(
"checkpoints/BindMPNN_v_32_010_25.pt",
map_location="cpu",
weights_only=True,
)
# Instantiate the upstream ProteinMPNN with model_type="ligand_mpnn",
# k_neighbors=checkpoint["num_edges"], and
# atom_context_num=checkpoint["atom_context_num"], then:
# model.load_state_dict(checkpoint["model_state_dict"], strict=True)
```
## ProteinMPNN-compatible checkpoints
These additional BindMPNN checkpoints use the **ProteinMPNN architecture** and can be loaded through the `soluble_mpnn` or `protein_mpnn` model options in the official LigandMPNN repository.
| Checkpoint | Backbone noise level | Protein neighbors |
|---|---:|---:|
| `BindMPNN_soluble_v_48_010.pt` | 0.10 Å | 48 |
| `BindMPNN_soluble_v_48_020.pt` | 0.20 Å | 48 |
Download them with:
```python
from huggingface_hub import hf_hub_download
for filename in ("BindMPNN_soluble_v_48_010.pt", "BindMPNN_soluble_v_48_020.pt"):
hf_hub_download(
repo_id="MurrellLab/SeqGivenStruct",
filename=filename,
local_dir="checkpoints",
)
```
Using the same installation described above, redesign chain B while keeping the other chains fixed:
```bash
python run.py \
--model_type soluble_mpnn \
--checkpoint_soluble_mpnn checkpoints/BindMPNN_soluble_v_48_010.pt \
--pdb_path complex.pdb \
--chains_to_design B \
--temperature 0.1 \
--seed 123 \
--batch_size 1 \
--number_of_batches 8 \
--out_folder outputs/bindmpnn_soluble
```
Replace the checkpoint filename with `BindMPNN_soluble_v_48_020.pt` for the 0.20 Å variant. Alternatively, use `--model_type protein_mpnn` with `--checkpoint_protein_mpnn` pointing to the same file.
These models use protein backbone coordinates and fixed-chain amino-acid identities. They do **not** accept target side-chain or nonprotein atom context; `--ligand_mpnn_use_side_chain_context` applies only to the LigandMPNN-compatible checkpoints above. Noise labels identify checkpoint variants and are independent of sampling temperature.
For programmatic loading, use `torch.load(path, map_location="cpu", weights_only=True)`, instantiate the upstream `ProteinMPNN` with `model_type="soluble_mpnn"` and `k_neighbors=checkpoint["num_edges"]`, then load `checkpoint["model_state_dict"]` strictly. The files contain only inference weights, `num_edges`, and `noise_level`.
## Attribution
Please acknowledge the corresponding upstream implementations: [LigandMPNN](https://github.com/dauparas/LigandMPNN) and [ProteinMPNN](https://github.com/dauparas/ProteinMPNN).
|