LigandMPNN

What is LigandMPNN?

LigandMPNN is a protein inverse-folding model for sequence design around non-protein atoms. It extends the ProteinMPNN family with context from ligands, metal ions, nucleotides, cofactors, and selected fixed side chains. The method was developed by Dauparas, Lee, and colleagues and published in Nature Methods in 2025.

The model changes amino acid sequence while keeping the submitted backbone coordinates fixed. Ligand coordinates must already be positioned in the same PDB as the protein. LigandMPNN does not dock a ligand or generate a new backbone.

How LigandMPNN works

LigandMPNN encodes the protein backbone together with nearby non-protein atoms, then samples amino acids autoregressively at redesigned positions. Fixed positions retain their submitted residues and can optionally contribute full side-chain coordinates as context.

Four official checkpoints are available. Their names describe the coordinate noise used during training:

Turning off Use atom context masks the parsed ligand atoms but still runs the selected LigandMPNN checkpoint. It does not switch the job to ProteinMPNN.

How to use LigandMPNN online

Upload a PDB or ENT structure, or fetch a PDB entry from RCSB, then choose sequence design or native-sequence scoring. ProteinIQ returns designed sequences, confidence and recovery statistics, downloadable FASTA and PDB files, and optional full-atom structures from the native side-chain packer.

Input

Protein and ligand files cannot be submitted separately. A separately prepared ligand must first be placed in the intended binding geometry and combined with the protein PDB.

Core settings

Scoring settings

Scoring returns the mean and standard deviation for all 20 amino acids at every parsed residue. It does not generate designed sequences.

Position and composition controls

| Setting | Description | | ------------------------------------ | ------------------------------------------------------------------------------------------------ | --------- | | Chains to design | Comma-separated chain IDs to redesign. Other parsed chains remain fixed. | | Fixed positions | PDB residue IDs, ranges, or whole chains to preserve, such as A15,A20-25,B. | | Redesigned positions | Residues to redesign while fixing everything else. It cannot be combined with Fixed positions. | | Parse chains only | Comma-separated chains to include from the structure. | | Homo-oligomer | Ties equivalent positions across equal-length chains. | | Symmetric residue groups | Ties arbitrary positions with source syntax such as A12,B12 \| A13,B13. | | Symmetry weights | Optional weights corresponding to each symmetric group. | | Exclude amino acids | Uppercase one-letter amino acid codes to omit globally, such as CW. | | Amino acid biases | Adds unrestricted global sampling biases for individual amino acids. | | Per-position amino acid biases | JSON mapping PDB residue IDs to residue-specific biases, such as {"A12":{"G":1.1}}. | | Per-position amino acid exclusions | JSON mapping PDB residue IDs to excluded amino acids, such as {"A12":"PG"}. |

Fixed positions, Redesigned positions, symmetry groups, and per-position constraints use the residue numbers and insertion codes from the submitted PDB.

Side-chain packing

Packing runs after sequence design and can return several independently packed structures for every design.

ProteinIQ caps projected native artifacts at 100 MiB per job. For large structures, reduce the number of sequences or packs per design, or reduce scoring passes and scores per pass, so every native result can be returned safely.

Interpreting LigandMPNN results

Design table

Confidence is model likelihood, not an experimental binding-affinity or activity prediction. Sequence recovery measures agreement with the submitted sequence and is not a success score.

Files

• The native FASTA preserves the model's headers and chain separators.
• The generated-design FASTA contains only sampled sequences.
• One backbone PDB is returned for every sampled sequence.
• Optional packed PDB files contain full-atom side-chain coordinates.
• Native statistics and run provenance remain downloadable.
• Scoring jobs return native PyTorch tensors plus a CSV with all per-position probabilities and deviations.

When to use LigandMPNN vs alternatives

LigandMPNN is the appropriate MPNN-family model when sequence choices should depend on an already positioned ligand, cofactor, metal, or nucleotide. ProteinMPNN is a simpler choice for backbone-only design. SolubleMPNN uses a ProteinMPNN-family checkpoint trained only on structures from soluble-protein PDB IDs and does not use ligand atoms during sequence design.

For workflows that need a new backbone, generate or predict the structure before sequence design. For example, RFdiffusion 2 can generate a backbone, while LigandMPNN assigns sequences to a supplied backbone and binding context.

LigandMPNN does not dock ligands, relax backbone coordinates, predict binding affinity, or guarantee function. Designed sequences and packed structures require structural review and experimental validation.