ProteinMPNN

ProteinMPNN represents the protein backbone as a k-nearest-neighbors graph where nodes are residues and edges encode spatial proximity. The model uses an encoder-decoder message-passing architecture:

• Encoder: Processes the full backbone structure bidirectionally, learning local and global geometric features
• Decoder: Autoregressively generates the amino acid sequence, conditioning each residue on the structure and previously decoded positions
• Edge features: Encode inter-residue distances, relative orientations, and backbone dihedral angles

This graph-based approach ensures the model respects the local chemical environment of each position while maintaining global structural consistency.

ProteinMPNN encodes the backbone using invariant spatial features that remain stable under rotation and translation:

• Pairwise distance matrices: Capture the Euclidean distances between Cα atoms of neighboring residues
• Relative orientation vectors: Encode the directional relationships between residue frames using dihedral and planar angles
• Virtual dihedral angles: Represent the torsion between consecutive Cα–Cα vectors

By operating on these SE(3)-invariant features rather than raw 3D coordinates, ProteinMPNN avoids the computational overhead of full equivariant networks while still producing sequences that are geometrically faithful to the input backbone. This design choice enables fast inference on large protein complexes.

A critical innovation in ProteinMPNN is the use of Gaussian noise augmentation on input backbone coordinates during training:

• Backbone coordinates are perturbed with small amounts of Gaussian noise (σ ≈ 0.1–0.3 Å)
• The model learns to generate sequences that are robust to minor structural fluctuations, mimicking the dynamic nature of real proteins
• This produces designs with higher experimental success rates because the sequences tolerate the slight backbone adjustments that occur during folding

In practice, this means ProteinMPNN-designed sequences are more likely to express solubly and fold correctly compared to earlier inverse folding tools like Rosetta's fixed-backbone design. The noise schedule can be tuned for specific design objectives—higher noise for flexible loops, lower noise for rigid active sites.

ProteinMPNN uses a tied encoder-decoder architecture where the same structural features condition every decoding step. Sequence generation is controlled by:

• Temperature parameter (T): Controls the diversity of generated sequences. Low T (0.1) produces near-deterministic, high-confidence designs; high T (1.0) explores diverse sequence space
• Autoregressive masking: The decoder predicts each residue sequentially, conditioning on previously sampled positions to maintain sequence-structure compatibility
• Multi-chain design: A single forward pass can simultaneously design sequences for all chains in a complex, respecting inter-chain packing constraints

This sampling flexibility allows users to balance between sequence recovery (matching native sequences) and sequence diversity (exploring novel functional variants) for a given backbone.

ProteinMPNN supports conditional design where specific residues or structural motifs are held fixed while the rest of the sequence is optimized:

• Catalytic site preservation: Active-site residues (e.g., catalytic triads) can be locked to maintain enzymatic function
• Binding interface design: Only surface-exposed positions are redesigned while the hydrophobic core remains fixed
• Epitope scaffolding: Known antibody-binding motifs are preserved while the surrounding scaffold is optimized for stability
• Metal-binding site retention: Cysteine and histidine residues coordinating zinc or iron are held constant

This is implemented through a residue-level mask that prevents the decoder from modifying specified positions, enabling function-preserving redesign of natural proteins for improved stability, solubility, or manufacturability.

ProteinMPNN is the canonical sequence design partner for structure diffusion models like RFdiffusion and FrameBuilder. The typical de novo design pipeline:

• Step 1: A diffusion model generates a novel protein backbone structure from random noise, often conditioned on a target binding site or symmetry specification
• Step 2: ProteinMPNN generates multiple candidate sequences for the backbone, sampling at different temperatures
• Step 3: Sequences are filtered by AlphaFold2 self-consistency—designs where AlphaFold2 predicts a structure matching the input backbone (RMSD < 1.0 Å) are selected for experimental validation

This structure-first, sequence-second paradigm has produced experimentally validated binders, enzymes, and self-assembling nanomaterials with no natural sequence homologs.

The computational task of designing an amino acid sequence that will fold into a specified target protein backbone structure. Unlike forward folding (structure prediction), inverse folding solves the sequence-design problem given a fixed 3D scaffold.

• ProteinMPNN is the current state-of-the-art method for this task
• Uses message-passing neural networks on protein graphs
• Achieves ~52% native sequence recovery on single-chain proteins
• Outperforms physics-based methods like Rosetta in both speed and accuracy

The computational creation of entirely new protein sequences and structures that do not exist in nature, designed to perform a specific function. ProteinMPNN is a cornerstone tool in this field.

• Backbone generation: Tools like RFdiffusion or FrameBuilder create novel scaffolds
• Sequence design: ProteinMPNN generates sequences predicted to fold into the scaffold
• Validation: AlphaFold or Rosetta checks if the designed sequence folds correctly
• Experimental testing: Genes are synthesized, proteins expressed, and structures solved via X-ray crystallography or cryo-EM

A class of generative models that learn to create novel protein structures by iteratively denoising random 3D coordinates, analogous to image generation techniques like DALL-E. These models generate the backbone scaffolds that ProteinMPNN then sequences.

• RFdiffusion: RoseTTAFold-based diffusion for unconditional and conditional backbone generation
• FrameDiff: SE(3) equivariant diffusion on residue frames
• Genie: SE(3)-equivariant diffusion for motif-scaffolding
• ProteinMPNN is the standard downstream sequence designer for all diffusion-generated backbones

A mathematical property of a neural network ensuring that its predictions transform consistently with the rotation and translation of the input 3D coordinates. This is critical for protein modeling because physical properties must be independent of the protein's orientation in space.

• ProteinMPNN achieves SE(3) invariance by operating on pairwise distances and local coordinate frames
• Equivariant architectures preserve directional information through network layers
• Key for generating physically realistic protein backbones in diffusion models
• Contrast with invariance, where outputs remain unchanged under transformations

The computational process of predicting the optimal 3D conformations of amino acid side chains onto a fixed protein backbone. ProteinMPNN designs the backbone sequence; side-chain packing determines the precise atomic positions.

• Uses discrete rotamer libraries derived from the Protein Data Bank
• Rosetta Packer is the most widely used tool for this task
• Critical for evaluating steric clashes and hydrogen bonding networks
• Often performed after ProteinMPNN sequence design to validate structural compatibility