Use LigandMPNN Online
Commercially Available LigandMPNN No-Code Web Server
Try LigandMPNN
LigandMPNN: Atomic Context-Conditioned Protein Sequence Design
LigandMPNN is a powerful deep learning framework designed to solve the protein sequence design problem by explicitly modeling non-protein atomic contexts. Developed by the Baker Lab at the University of Washington, it generalizes the state-of-the-art ProteinMPNN architecture to incorporate small molecules, nucleotides, metals, and other non-protein components.
By reasoning about the full biomolecular system, LigandMPNN enables the precise design of enzymes, biosensors, and binders that require high chemical complementarity between a protein and its target.
Key Innovations: Explicit Non-Protein Modeling
LigandMPNN overcomes the critical limitation of previous models that only considered protein backbone coordinates.
Atomic Contextual Graphs: Operates on three distinct graphs: a protein-only graph, an intra-ligand graph (element types and inter-atomic distances), and a protein-ligand graph that transfers information from ligand atoms to protein residues.
Message Passing for Richer Data: Intra-ligand message passing increases the richness of information transferred to the protein, allowing the model to "see" specific chemical environments.
Sidechain Conformation Generation: Unlike models that only output sequences, LigandMPNN predicts sidechain torsion angles (x1–x4), facilitating detailed evaluation of binding interactions.
Fast and Scalable Architecture: The neural network features 2.62 million parameters and remains highly efficient, designing sequences for a 100-residue protein in approximately 0.9 seconds on a single CPU.
Generalization Mastery: Due to chemical similarities (e.g., carbon, oxygen, nitrogen), the model can generalize from sidechain atoms to novel small-molecule contexts even without specific training data.
Performance Benchmarks: Superior Recovery Accuracy
LigandMPNN significantly outperforms both physics-based methods (Rosetta) and predecessor models (ProteinMPNN) in native sequence recovery for residues interacting with non-protein atoms.
Context Type | Rosetta Recovery | ProteinMPNN Recovery | LigandMPNN Result | Key Finding |
Small Molecules | 50.4% | 50.5% | 63.3% | Significant jump in binding site accuracy |
Nucleotides | 35.2% | 34.0% | 50.5% | Overcomes large-scale atom modeling challenges |
Metals | 36.0% | 40.6% | 77.5% | Near-perfect recovery of metal coordination sites |
Sidechain Packing | 76.0% X1 | 83.3% X1 | 86.1% X1 | Higher recovery of native torsion angles |
Scientific Breakthroughs in Interaction Design
Rescuing Non-Functional Designs
Experimental characterization has shown that LigandMPNN can "rescue" weak or non-binding proteins designed by older methods. For example, LigandMPNN redesigns of rocuronium and cholic acid binders successfully introduced new sidechain-ligand hydrogen bonds that were missing in original Rosetta designs.
Enzyme and DNA Binder Engineering
LigandMPNN has been experimentally validated across more than 100 protein-DNA and protein-small molecule interfaces. It has successfully generated sequence-specific DNA-binding proteins that recognize targets in the major groove, with crystal structures (e.g., PDB ID 8TAC) matching the design models with high fidelity.
Multi-State and Symmetric Design
Utilizing a random autoregressive decoding scheme, the model facilitates the design of complex architectures, including symmetric protein assemblies and multi-state hinge proteins.
LigandMPNN on Tamarind Bio: Comprehensive Atomic Design
Tamarind Bio provides an optimized, no-code environment to deploy LigandMPNN for high-throughput design of enzymes, sensors, and therapeutic binders.
Integrated Non-Protein Handling: Seamlessly manage ligands, ions, and nucleic acids in your design loop without manual force field parameterization.
Interactive Confidence Reports: Leverage per-residue predicted confidence scores that correlate strongly with actual sequence recovery accuracy to identify the most promising leads.
How to Use LigandMPNN on Tamarind Bio
Access the Toolkit: Log in to tamarind.bio and select the LigandMPNN tool.
Upload Molecular Complex: Provide a PDB file containing both the protein backbone and the atomic context (small molecule, ion, or nucleotide).
Define Designable Residues: Use the interactive 3D viewer to select which residues should be redesigned (typically those within 5.0 Å of the ligand).
Set Modeling Constraints: Optionally input specific sidechain coordinates to stabilize functional sites of interest.
Run Design Trajectories: The platform executes message-passing cycles to generate optimized sequences and their corresponding sidechain conformations.
Evaluate & Validate: Review sequence logos and confidence maps, then download top candidates for experimental synthesis.