lamm-mit/skills/ligandmpnn

LigandMPNN Ligand-Aware Sequence Design

Extends ProteinMPNN to design sequences around small molecules, metal ions, nucleic acids, and other non-protein entities. Essential for enzyme active site design and ligand-binding protein engineering.

Installation

git clone https://github.com/dauparas/LigandMPNN
cd LigandMPNN
pip install -e .
# Download model weights
bash get_model_params.sh

Key Difference from ProteinMPNN

ProteinMPNN only sees protein backbone atoms. LigandMPNN encodes ligand atoms as additional context nodes, so designed sequences are informed by the chemical environment of the binding site.

Basic Usage

Design Around Small Molecule

python3 run.py \
    --model_type "ligand_mpnn" \
    --checkpoint_ligand_mpnn "model_params/ligandmpnn_v_32_010_25.pt" \
    --pdb_path complex.pdb \
    --out_folder output/ \
    --number_of_batches 8 \
    --batch_size 4

Fix Non-Binding Residues (Design Interface Only)

python3 run.py \
    --model_type "ligand_mpnn" \
    --checkpoint_ligand_mpnn "model_params/ligandmpnn_v_32_010_25.pt" \
    --pdb_path enzyme.pdb \
    --out_folder output/ \
    --redesigned_residues "A42 A67 A89 A134 A156"  # active site only

Metal Coordination Design

python3 run.py \
    --model_type "ligand_mpnn" \
    --checkpoint_ligand_mpnn "model_params/ligandmpnn_v_32_010_25.pt" \
    --pdb_path metal_complex.pdb \
    --out_folder output/ \
    --use_side_chain_context 1  # include side-chain atoms in context

Protein-DNA/RNA Interface Design

python3 run.py \
    --model_type "per_residue_label_membrane_mpnn" \
    --checkpoint_path "model_params/ligandmpnn_v_32_020_25.pt" \
    --pdb_path dna_complex.pdb \
    --out_folder output/ \
    --chains_to_design "A"  # design protein chain A around DNA

Model Types

| Model | Use Case | |-------|---------| | ligand_mpnn | Small molecules, metal ions | | per_residue_label_membrane_mpnn | Membrane proteins (hydrophobic label per residue) | | global_label_membrane_mpnn | Transmembrane topology-aware | | soluble_mpnn | See solublempnn skill |

Python API

import subprocess
import json

def design_with_ligand(pdb_path, out_folder, n_seqs=32, temp=0.1):
    cmd = [
        "python3", "LigandMPNN/run.py",
        "--model_type", "ligand_mpnn",
        "--checkpoint_ligand_mpnn", "model_params/ligandmpnn_v_32_010_25.pt",
        "--pdb_path", pdb_path,
        "--out_folder", out_folder,
        "--number_of_batches", str(n_seqs // 4),
        "--batch_size", "4",
        "--temperature", str(temp),
    ]
    result = subprocess.run(cmd, capture_output=True, text=True)
    if result.returncode != 0:
        raise RuntimeError(result.stderr)
    return out_folder

Output

Same FASTA format as ProteinMPNN with additional ligand-context score:

>design_0, score=0.876, ligand_score=0.234, seq_recovery=0.51
MTEYKLVVVGAGGVGKSALTIQLIQNHFVDEYDPTIEDSY...

Lower ligand_score = better accommodation of ligand.

When to Use LigandMPNN vs ProteinMPNN

| Scenario | Use | |----------|-----| | No ligand in design | ProteinMPNN | | Small molecule binding site | LigandMPNN | | Metal coordination | LigandMPNN | | DNA/RNA interface | LigandMPNN | | Membrane protein | LigandMPNN (membrane model) | | Maximize solubility | SolubleMPNN |

Typical Workflow for Enzyme Design

1. Generate backbone with RFdiffusion (motif scaffolding of active site)
2. Place ligand in active site (docking or manual placement)
3. Run LigandMPNN with --redesigned_residues = active site shell
4. Generate 200–500 sequences
5. Validate with structure prediction (Chai/Boltz) + docking (DiffDock)
6. Filter by pLDDT > 75 and predicted ligand contacts