Use ProteinMPNN Online
Commercially Available ProteinMPNN No-Code Web Server
ProteinMPNN: Fast Deep Learning Protein Sequence Design from Backbone Structures
ProteinMPNN is a message passing neural network for protein sequence design: given a backbone structure, it predicts amino acid sequences likely to fold into it. It predicts sequences autoregressively with an order-agnostic decoder, and the text states that the Institute for Protein Design adopted it as its standard sequence design approach.
Designing a sequence for a target backbone is the key step after generating new structures, whether by hallucination, diffusion or manual design. The authors report that physically based Rosetta sequence design is slow and often yields sequences that do not express or fold, and that sequences from structure-generation methods can be insoluble. ProteinMPNN is aimed at fixing both speed and success rate.
Attribute | Details |
|---|---|
Task | Backbone-conditioned protein sequence design |
Architecture | Message passing network, 3 encoder and 3 decoder layers in the baseline |
Decoding | Order-agnostic, so fixed regions can be decoded first as context |
Training data | PDB assemblies, X-ray or cryo-EM better than 3.5 A, clustered at 30% identity (25,361 clusters) |
Native sequence recovery | 52.4% vs 32.9% for Rosetta on 402 test monomer backbones |
Speed | 1.2 s vs 4.3 min for a 100-residue protein |
How ProteinMPNN Works: Order-Agnostic Decoding, Symmetry and Multistate Design
Richer inputs: distances between N, C-alpha, C, O and a virtual C-beta atom plus edge updates raised native monomer recovery from 41.2% to 50.5% in ablations.
Flexible decoding order: the order is randomly sampled in training and chosen at inference, which lets users fix some positions and design the rest.
Multi-chain design: relative positional encoding and a same-chain versus different-chain feature handle complexes.
Tied positions: logits for symmetry-equivalent or repeated positions are averaged, supporting homo-oligomers and repeat proteins. Averaging logits gave 55% recovery on homomers versus 52% with no symmetry.
Multistate design: averaged or weighted logits across states allow positive and negative design.
Sampling temperature and noise: temperature raises diversity at small recovery cost, and average log probability can rank designs.
Experimental Validation of ProteinMPNN Designs
Test | Result |
|---|---|
Hallucinated backbones, solubility | 73 of 96 ProteinMPNN designs soluble (median 247 mg/L) vs median 9 mg/L for original sequences |
Cyclic C5/C6 oligomers, correct oligomeric state | 27.7% for ProteinMPNN vs 0% for Rosetta |
Tetrahedral nanoparticles | 13 of 76 sequences formed assemblies of the expected size |
Crystal structures | About 2.35 A RMSD over 130 residues (monomer) and 1.2 A over 232 residues (repeat protein) |
ProteinMPNN also redesigned Rosetta-designed scaffolds so that most were confidently predicted to fold by AlphaFold, and designed scaffolds showed strong binding to Grb2 SH3 by biolayer interferometry.
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 ProteinMPNN on Tamarind Bio
Sequences for generated backbones: assign sequences to backbones from hallucination or other generative methods.
Symmetric assemblies: design homo-oligomers, cyclic oligomers and nanoparticles using tied positions.
Partial redesign: keep functional residues fixed and redesign the rest.
Binder scaffolds: design scaffolds for binding a target, as shown for Grb2 SH3.
How to Use ProteinMPNN on Tamarind Bio
Open the tool: log in to tamarind.bio and open ProteinMPNN.
Upload the backbone: provide a PDB structure of the target backbone, which may be a monomer or complex.
Fix positions: mark residues to keep unchanged, if exposed.
Tie positions: specify symmetry-equivalent positions for oligomers or repeats, if exposed.
Set sampling: choose the temperature and number of sequences, if exposed. Higher temperature increases diversity.
Run the job: submit and wait for completion.
Download and rank: download sequences and rank by average log probability if provided.
Check downstream: predict structures of the top sequences with a structure predictor before ordering genes.
Things to Keep in Mind
Native sequence recovery does not reliably predict folding, since a single substitution can block folding.
Experimental success rates are moderate in some categories, for example 50 of 96 monomers had the target oligomeric state.
Recovery is sensitive to crystallographic resolution.
The method handles proteins only; protein-nucleic acid and protein-small-molecule design are described as future extensions.
Source: ProteinMPNN preprint on bioRxiv, DOI 10.1101/2022.06.03.494563 (https://www.biorxiv.org/content/10.1101/2022.06.03.494563v1). No code repository was listed on the page read.