How to Use LigandMPNN

LigandMPNN is an inverse folding model that designs amino-acid sequences from 3D structure while conditioning on nearby atomic context. In contrast to ProteinMPNN, the model can use ligands, cofactors, nucleic acids, and other non-protein atoms as part of the design problem, which makes it especially useful for active-site redesign, ligand-binding proteins, and interface-focused engineering.

On Neurosnap, researchers upload an Input Structure, mark Fixed Positions they want to preserve, and then control exploration with Redesign Selection, Number Sequences, Sampling Temperature, and Model Type. The service also exposes ProteinMPNN and SolubleMPNN variants, so one workflow can cover ligand-conditioned design, backbone-conditioned redesign, and solubility-oriented sequence generation.

This makes LigandMPNN practical for enzyme engineering, binder optimization, and functional homolog design when a backbone or complex structure is already known but the sequence still needs to be reworked.

LigandMPNN extends the message-passing inverse-folding framework by including atomic context outside the protein backbone. The scientific consequence is important: residues are no longer chosen only to satisfy fold geometry, but also to accommodate nearby chemical groups, pocket shape, and interface context. That is why the method is more appropriate than sequence-only design for ligand-binding or cofactor-dependent proteins.

On Neurosnap, Fixed Positions and Redesign Selection define the designable region. Researchers can freeze catalytic residues, preserve a binding epitope, or invert the selection to redesign only a narrow patch. Model Type lets users choose LigandMPNN, classic ProteinMPNN, or SolubleMPNN with different noise settings, which changes how conservative or exploratory the design process will be.

The results are meant for triage, not just generation. Sequence-level confidence, ligand-context confidence, sequence recovery, residue-probability heatmaps, and per-position uncertainty help users decide whether a proposed design is consensus-like, chemically well supported, or worth deeper experimental diversification.

• Atomic-context design: LigandMPNN can account for ligands, cofactors, and other nearby atoms instead of designing from backbone geometry alone.
• Precise region control: Fixed Positions and Redesign Selection make it easy to preserve catalytic or binding-critical residues while redesigning the surrounding context.
• Multiple model families: LigandMPNN, ProteinMPNN, and SolubleMPNN variants let researchers match the design strategy to the problem.
• Residue-level interpretation: Confidence summaries and amino-acid probability maps help distinguish consensus-like designs from uncertain exploratory ones.