Use LigandMPNN Online
Commercially Available LigandMPNN No-Code Web Server
LigandMPNN: Atomic Context-Conditioned Protein Sequence Design for Ligands, Nucleotides and Metals
LigandMPNN is a deep-learning method for designing protein sequences from a fixed backbone while modeling non-protein atoms such as small molecules, nucleotides and metal ions. It was developed by Justas Dauparas, Gyu Rie Lee, Robert Pecoraro, Linna An, Ivan Anishchenko, Cameron Glasscock and David Baker at the University of Washington (Institute for Protein Design and Howard Hughes Medical Institute), and is described in a bioRxiv preprint. Code is available at github.com/dauparas/LigandMPNN.
ProteinMPNN designs sequences from protein backbone coordinates alone. Many functional proteins, including enzymes, sensors and binders, depend on interactions with ligands, DNA or metals, which a protein-only model cannot see.
LigandMPNN adds that atomic context, and can also generate side-chain conformations to help assess binding interactions.
Item | Detail from the preprint |
|---|---|
Inputs | Backbone coordinates plus non-protein context atoms; optional side-chain context |
Parameters | About 2.62 million (ProteinMPNN: 1.66 million) |
CPU speed | About 0.9 s per 100 residues (ProteinMPNN: 0.6 s) |
Training data | PDB assemblies (to Dec 2022) at better than 3.5 A resolution |
Test sets | 317 small-molecule, 74 nucleotide, 83 metal structures |
How LigandMPNN Models Atomic Context
Protein graph: residues are nodes linked to nearest neighbors, with backbone geometry encoded as distances among five backbone atom types.
Protein-ligand graph: the 25 closest ligand atoms per residue are used, with one-hot element types for ligand nodes and interatomic distances for edges.
Random autoregressive decoding: this supports symmetric and multi-state designs.
Side-chain packing: the same architecture predicts chi1 to chi4 torsion angles, decoding chi1 first and feeding each sampled angle into the next.
Noise during training: 0.1 A Gaussian noise limits memorization of the backbone.
Sequence Recovery and Side-Chain Packing Results
Native sequence recovery near ligands
Measured on residues with side-chain atoms within 5.0 A of context atoms (median over ten designs per protein), LigandMPNN outperformed both Rosetta and ProteinMPNN.
Context | Rosetta | ProteinMPNN | LigandMPNN |
|---|---|---|---|
Small molecules | 50.4% | 50.4% | 63.3% |
Nucleotides | 35.2% | 34.0% | 50.5% |
Metals | 36.0% | 40.6% | 77.5% |
Side chains and experiments
For chi1 accuracy within 10 degrees on small-molecule sites, LigandMPNN scored 86.1% versus 83.3% for ProteinMPNN and 76.0% for Rosetta, and the authors report it runs about 250 times faster than Rosetta. Experimentally, LigandMPNN redesign rescued binding of rocuronium in a Rosetta design with no observed binding, and improved cholic acid binding from a weak starting design.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with LigandMPNN on Tamarind Bio
Small-molecule binders and sensors: design new binders or improve weak ones.
DNA-binding proteins: design sequence-specific protein-DNA interfaces; a crystal structure of a LigandMPNN-designed binder recapitulated the model.
Iterative binder design: the authors cite use across scaffold families for ligands such as apixaban, methotrexate, thyroxine and heme.
Enzyme design: cited by the authors as a motivating application.
How to Use LigandMPNN on Tamarind Bio
Open the tool: log in to tamarind.bio and open LigandMPNN.
Upload a structure: provide a backbone that includes the ligand, nucleotide or metal atoms you want to design around.
Select residues: choose which positions to design or fix, if exposed.
Set options: pick the number of designs, and noise or side-chain context settings, if exposed.
Run the job: submit and wait for designed sequences.
Download outputs: retrieve sequences and, if exposed, packed side-chain structures.
Evaluate: assess designs with structure prediction or physics-based scoring, then test experimentally.
Things to Keep in Mind
Use caution with compounds containing elements rarely or never seen in the PDB; the authors suggest mapping to the closest common element.
Chi3 and chi4 predictions remain poor for all methods tested.
Side-chain packing gains over ProteinMPNN are marginal.
Experimental validation in this preprint covers only a small set of examples.
Recovery depends on the training noise level, which must be matched when comparing models.
Source: Dauparas J, Lee GR, Pecoraro R, An L, Anishchenko I, Glasscock C, Baker D. Atomic context-conditioned protein sequence design using LigandMPNN. bioRxiv preprint, 2023 (v1). DOI: 10.1101/2023.12.22.573103. Preprint: biorxiv.org/content/10.1101/2023.12.22.573103v1. Code: github.com/dauparas/LigandMPNN.