ProteinMPNN

ProteinMPNN represents the protein backbone as a k-nearest-neighbors graph where nodes are residues and edges encode spatial relationships between Cα atoms. The model performs iterative message passing to learn local structural environments:

• Node features: Backbone dihedral angles (φ, ψ, ω) and pairwise distance/orientation encodings
• Edge features: Inter-residue distances, relative orientations, and chain identity information
• Encoder-decoder architecture: An encoder processes the full backbone graph, then an autoregressive decoder generates sequences residue-by-residue

This graph-based approach ensures the model captures both local steric constraints and long-range packing interactions critical for sequence design.

A defining innovation of ProteinMPNN is training on backbone coordinates with Gaussian noise added to atomic positions. This teaches the model to design sequences that fold robustly even when the input structure contains small errors or conformational flexibility:

• Noise scale: Typically 0.1–0.3 Å standard deviation applied to Cα coordinates during training
• Effect: Prevents overfitting to exact backbone geometries; sequences remain foldable across slight structural variations
• Practical benefit: Enables successful design from NMR ensembles, cryo-EM structures at moderate resolution, or computationally predicted backbones with inherent uncertainty

This noise robustness is a key reason ProteinMPNN outperforms physics-based design methods like Rosetta on experimental validation benchmarks.

ProteinMPNN generates sequences using an autoregressive decoder that predicts each residue's amino acid identity conditioned on the backbone and previously decoded residues. A critical training trick is decoding order randomization:

• During training, the order in which residues are decoded is randomly permuted for each example
• This forces the model to learn permutation-invariant sequence distributions rather than relying on N-to-C terminal biases
• At inference, any decoding order can be used, enabling constrained design where specific positions are fixed and the model adapts surrounding residues accordingly

The autoregressive formulation naturally captures epistatic interactions between residue choices, producing sequences with higher folding stability than independent-per-position classifiers.

ProteinMPNN natively handles multi-chain complexes and symmetric assemblies through specialized architectural features:

• Chain identity encoding: A one-hot vector appended to node features distinguishes residues belonging to different chains
• Tied decoding for symmetry: For symmetric complexes (e.g., cyclic, dihedral, icosahedral), the model can tie sequence predictions across symmetry-related positions, ensuring identical subunits
• Interface-aware design: The message-passing graph includes inter-chain edges, allowing the model to optimize interface packing and hydrogen bonding between binding partners

This capability has enabled the computational design of protein nanocages, symmetric signaling complexes, and de novo binders with experimental validation rates far exceeding previous methods.

ProteinMPNN supports temperature-controlled sampling during autoregressive decoding, enabling users to tune the diversity of generated sequences:

• Low temperature (T → 0): Produces high-confidence, near-consensus sequences with maximal predicted recovery rates
• High temperature (T > 1): Generates diverse sequence variants exploring alternative packing solutions and surface compositions
• Fixed-position constraints: Specific residues can be held constant while the model designs around them, useful for preserving catalytic sites or binding motifs

This sampling flexibility makes ProteinMPNN valuable for library design in directed evolution campaigns, where diverse but foldable variants are screened for improved function.