ProteinMPNN

ProteinMPNN constructs a k-nearest-neighbors graph from the Cα coordinates of the input backbone, where edges connect spatially proximal residues regardless of sequence distance. The encoder processes this graph using message-passing layers that update node and edge features iteratively.

• Node features encode local backbone geometry: dihedral angles (φ, ψ, ω), pairwise distances, and relative orientations
• Edge features capture inter-residue distances and relative coordinate frames
• Decoder autoregressively predicts amino acid probabilities at each position, conditioned on the encoded representations and previously decoded residues
• The k=48 default neighborhood size ensures sufficient context for each residue's structural environment

This graph-based approach captures both local and non-local spatial interactions critical for fold stability.

ProteinMPNN is trained with Gaussian noise added to the input backbone coordinates, forcing the model to learn sequence-structure relationships that are robust to structural perturbations. This is critical for real-world applications where backbone models contain errors.

• Training adds 0.1–0.3 Å standard deviation noise to Cα positions
• The model learns to design sequences that fold correctly even when the input structure is imperfect
• This noise augmentation mimics the resolution limitations of cryo-EM density maps and NMR ensembles
• Sequences designed from noisy backbones still achieve high experimental expression and solubility rates

The noise tolerance makes ProteinMPNN particularly valuable for designing sequences into cryo-EM models that may have local coordinate errors.

The decoder generates amino acid sequences autoregressively, predicting one residue at a time while conditioning on previously decoded positions. A key innovation is tied decoding, where the same decoder weights are applied iteratively across all sequence positions.

• Decoding order is randomized during training to prevent positional bias
• At inference, multiple decoding orders can be sampled to generate diverse sequence candidates
• The model outputs a probability distribution over 20 amino acids at each position
• Temperature sampling controls sequence diversity: lower temperatures yield higher-confidence designs, higher temperatures explore more sequence space
• Greedy decoding selects the single most probable amino acid at each position for deterministic design

This autoregressive framework captures higher-order amino acid correlations that independent site predictions miss.

The encoder transforms raw backbone coordinates into a rich geometric representation using invariant and equivariant features that capture the local structural environment of each residue.

• Dihedral angles (φ, ψ, ω) describe the local backbone conformation
• Pairwise distance matrices between Cα, C, and N atoms encode spatial relationships
• Relative orientation frames capture the rotational relationship between residue coordinate systems
• These features are SE(3)-invariant, meaning they do not change under global rotation or translation of the structure
• The encoder uses multiple message-passing layers to propagate information across the KNN graph

This geometric encoding ensures the model understands the full 3D context of each residue position before sequence design begins.

ProteinMPNN is widely adopted in cryo-EM structure determination pipelines to validate and optimize atomic models built into density maps. The designed sequences serve as an orthogonal check on model quality.

• Sequences designed by ProteinMPNN into a cryo-EM model are compared to the native sequence; high recovery indicates a well-built backbone
• Low recovery at specific positions flags potential register shifts or modeling errors
• Designed sequences can be experimentally tested for expression and folding to validate structural hypotheses
• Integration with ModelAngelo and Coot enables iterative model building and sequence design cycles
• The method helps resolve ambiguities in side-chain placement where density is weak

This application bridges computational protein design with experimental structural biology, providing a powerful validation tool.

ProteinMPNN incorporates training strategies that bias designed sequences toward soluble, well-expressed proteins rather than just maximizing sequence recovery. This is achieved through dataset curation and loss function design.

• Training data is filtered to include only monomeric, soluble proteins from the PDB
• The model learns to avoid hydrophobic patches that would cause aggregation
• Designed sequences show higher experimental solubility than Rosetta-designed counterparts
• Surface residue predictions favor polar and charged amino acids appropriate for aqueous environments
• Core packing predictions maintain hydrophobic residues for thermodynamic stability

This bias toward biophysical realism makes ProteinMPNN designs more likely to succeed in experimental validation than purely structure-matching approaches.

The inverse of the classic protein folding problem: given a target 3D backbone structure, predict an amino acid sequence that will fold into it. ProteinMPNN solves this using a message-passing neural network that reasons over the spatial graph of the backbone, encoding geometric features like Cα-Cα distances and dihedral angles to autoregressively decode a sequence optimized for structural stability.

A critical validation metric for designed sequences: the designed sequence is fed into a structure prediction tool like AlphaFold2 or ESMFold. If the predicted structure closely matches the original target backbone (low RMSD), the design is considered self-consistent. ProteinMPNN designs achieve remarkably high self-consistency, indicating the network has internalized the sequence-structure mapping.

ProteinMPNN serves as an orthogonal validation tool for cryo-EM models. By redesigning sequences onto a model built into the density, researchers can:

• Identify register shifts: Incorrect backbone tracing leads to nonsensical designed sequences.
• Assess model quality: High native sequence recovery by ProteinMPNN correlates with well-built, geometrically sound models.
• Resolve ambiguous density: In regions of poor density, the network's confidence scores guide manual rebuilding efforts.

Beyond validation, ProteinMPNN is a generative design engine. Given a novel, computationally designed backbone—such as those from Rosetta or RFdiffusion—ProteinMPNN generates sequences predicted to fold into that structure. Key capabilities include:

• Fixed backbone design: Sequence generation for a rigid target.
• Flexible backbone design: Accommodating slight backbone noise or ensemble inputs.
• Conditional design: Fixing specific residue identities (e.g., catalytic site residues) while designing the remainder of the sequence.

A distinguishing feature of ProteinMPNN is its robustness to backbone noise. Trained with Gaussian noise added to input coordinates, the network generates viable sequences even from imperfect backbones—such as those derived from moderate-resolution cryo-EM maps or computationally generated scaffolds. This noise conditioning prevents overfitting to atomic-level details that may not be physically accurate in the target structure.