ProteinMPNN

ProteinMPNN represents the protein backbone as a k-nearest-neighbors graph where nodes are residues and edges encode spatial relationships. The model uses an encoder-decoder architecture with message-passing layers that iteratively update node features based on neighboring residue geometries.

• Encodes Cα-Cα distances, relative orientations, and backbone dihedral angles
• Uses invariant features (distances) rather than raw coordinates for rotational invariance
• Decoder autoregressively predicts amino acid probabilities at each position
• No multiple sequence alignment (MSA) required—operates on single backbone structures

ProteinMPNN is trained with Gaussian noise added to backbone coordinates, making it robust to structural perturbations. This noise augmentation teaches the model to recover viable sequences even when backbone geometries deviate from ideal crystallographic precision.

• Trained on the Protein Data Bank (PDB) with structures from X-ray crystallography and cryo-EM
• Noise levels up to 0.1 Å standard deviation on atomic coordinates
• Enables design from computationally predicted structures (AlphaFold, Rosetta) that may contain minor errors
• Produces sequences that fold correctly despite backbone uncertainty

ProteinMPNN employs a tied decoding strategy where the same encoder network processes the entire graph, and a lightweight decoder predicts residues in randomized order. This eliminates the sequential bias inherent in left-to-right autoregressive models.

• Residues are decoded in random permutations during training
• Model learns to condition on any subset of already-predicted residues
• Enables iterative refinement—resample problematic positions while keeping others fixed
• Supports conditional design where known functional residues are held constant

ProteinMPNN-generated sequences consistently exhibit superior solubility and thermostability compared to physically-based design methods like Rosetta. The model implicitly learns to avoid hydrophobic patches and aggregation-prone motifs.

• Designs show higher expression yields in E. coli and mammalian systems
• Predicted structures of designed sequences match target backbones with sub-Ångström RMSD
• Outperforms Rosetta in experimental success rates for de novo protein design
• Used to design soluble variants of membrane proteins and aggregation-prone scaffolds

ProteinMPNN natively handles multi-chain protein complexes and cyclic symmetries, making it suitable for designing oligomeric assemblies, nanocages, and repeat proteins.

• Accepts multiple chain inputs with inter-chain edge features
• Supports C2, C3, C4, and higher cyclic symmetries by tiling backbone coordinates
• Used to design protein nanocages for drug delivery and vaccine scaffolds
• Enables design of heteromeric complexes with specified binding interfaces

ProteinMPNN supports conditional sequence generation where specified residues or functional motifs are held fixed while the model designs the surrounding scaffold. This enables functional site grafting onto stable scaffolds.

• Constrain catalytic residues in enzyme active sites during design
• Fix binding interface residues while optimizing scaffold stability
• Design epitope-focused immunogens by holding antigenic regions constant
• Used to scaffold therapeutic peptide motifs onto stable protein frameworks

ProteinMPNN excels at designing sequences for completely novel backbone structures that have no natural homologs. Unlike physics-based methods, it implicitly learns to avoid internal cavities and unsatisfied hydrogen bonds.

• Generates sequences for hallucinated backbones from diffusion models like RFdiffusion
• Achieves high experimental success rates for designs with no evolutionary precedent
• Produces sequences that express solubly in E. coli at rates exceeding traditional Rosetta design

By conditioning on backbone structures at elevated temperatures or targeting rigidifying mutations, ProteinMPNN designs variants with enhanced thermal tolerance.

• Identifies surface loop regions where glycine substitutions reduce backbone flexibility
• Designs complementary hydrophobic cores that resist thermal unfolding
• Enables engineering of industrial enzymes stable at 60-80°C for biocatalysis applications

ProteinMPNN can redesign the non-contacting surfaces of protein binders while preserving the binding interface geometry, enabling affinity maturation and specificity engineering.

• Fixes paratope residues while redesigning the scaffold for improved stability
• Introduces solubility-enhancing mutations on exposed surfaces without disrupting binding
• Generates diverse sequence libraries for directed evolution campaigns starting from computational designs

The model natively handles cyclic and dihedral symmetry constraints, making it ideal for designing self-assembling protein nanomaterials.

• Designs sequences for C2 through C8 symmetric assemblies simultaneously
• Ensures consistent inter-subunit packing across all interfaces
• Enables creation of virus-like particles and protein cages for drug delivery

When physics-based design methods produce sequences that aggregate or fail to express, ProteinMPNN provides an orthogonal computational rescue strategy.

• Redesigns Rosetta-generated sequences that fail experimental validation
• Removes cryptic aggregation-prone regions while preserving fold specificity
• Recovers soluble expression for previously intractable computational designs
• Acts as a sequence optimization filter in multi-stage design pipelines

ProteinMPNN designs scaffolds around catalytic residues fixed in their catalytically competent geometry, enabling de novo enzyme creation.

• Preserves precise geometric arrangement of catalytic triads or metal-binding sites
• Designs the surrounding pocket to accommodate specific substrates
• Generates sequences for retro-aldolases, hydrolases, and other computationally designed enzymes
• Integrates with transition state docking for full active site specification

The fundamental computational task solved by ProteinMPNN. Given a target 3D backbone structure, the model predicts an amino acid sequence that will fold into that exact shape. This is the inverse of the classic protein folding problem.

• Input: 3D coordinates (N, Cα, C atoms)
• Output: A probability distribution over the 20 amino acids at each position
• Key Metric: Sequence Recovery Rate — the percentage of native residues correctly predicted

The standard benchmark metric for evaluating inverse folding models. It measures the percentage of positions where the model predicts the exact native amino acid found in the original protein.

• ProteinMPNN achieves 52.4% on single-chain proteins
• Higher than physics-based methods like Rosetta (~32%)
• A high rate indicates the model has learned the sequence-structure mapping encoded in evolution

A conceptual mapping of all possible protein sequences to their biological fitness or functional activity. ProteinMPNN enables efficient navigation of this landscape by generating sequences that are guaranteed to fold into a desired structure.

• Used to visualize evolutionary trajectories
• Guides protein engineering toward improved stability, binding, or catalysis
• ProteinMPNN's noise-augmented training makes it robust to small structural perturbations

The process of navigating a protein language model's latent space to generate novel sequences with altered properties. While ProteinMPNN operates on structural inputs, its learned representations can be combined with sequence-based models for controlled design.

• Involves interpolating or perturbing learned embeddings
• Enables property-conditioned sequence generation
• Bridges structure-based and sequence-based design paradigms

A high-throughput experimental method that assays the functional effect of thousands of single amino acid substitutions across a protein. The resulting data provides rich training and validation sets for models like ProteinMPNN.

• Generates variant effect maps for entire proteins
• Validates computational predictions of tolerated vs. destabilizing mutations
• ProteinMPNN's log-probability scores correlate with experimental stability measurements