Inverse Folding

The core computational task is fixed-backbone sequence design. The model takes a rigid 3D backbone scaffold (N, Cα, C coordinates) as input and outputs a probability distribution over the 20 canonical amino acids at each position. Unlike structure prediction, which searches conformational space, inverse folding searches sequence space constrained by a fixed geometry. The objective is to maximize P(sequence | backbone), ensuring the designed sequence has that backbone as its lowest-energy state.

Leading models like ProteinMPNN represent the protein backbone as a k-nearest-neighbors graph where nodes are residues and edges encode spatial relationships. Key architectural features include:

• Message-passing neural networks (MPNNs) that propagate information between spatially adjacent residues
• Encoder-decoder architectures where the encoder processes backbone geometry and the decoder autoregressively generates sequences
• Invariant feature construction using distances, dihedral angles, and relative spatial orientations rather than raw coordinates This graph-based formulation naturally captures the local packing interactions that determine sequence-structure compatibility.

A defining innovation in models like ProteinMPNN is training with Gaussian noise added to backbone coordinates. This teaches the model to design sequences that fold robustly even when the target backbone is slightly perturbed. Benefits include:

• Increased soluble expression in experimental validation
• Tolerance to backbone inaccuracies from structure prediction models
• Higher sequence recovery rates on native backbones
• Improved performance on flexible loops and surface residues The noise-augmentation strategy effectively encodes a smooth energy landscape around the target structure, producing sequences that are less brittle to minor structural variations.

Inverse folding models employ two primary decoding strategies:

• Autoregressive decoding: Residues are predicted sequentially (N-to-C terminus), with each prediction conditioned on previously designed residues. This captures epistatic coupling between sequence positions.
• One-shot decoding: All positions are predicted simultaneously in a single forward pass, offering faster inference but potentially missing cooperative interactions. ProteinMPNN uses an autoregressive decoder with a random decoding order during training, which allows it to generate diverse sequence samples for the same backbone—critical for exploring sequence space in design campaigns.

The primary evaluation metric for inverse folding is native sequence recovery rate—the percentage of positions where the model predicts the wild-type amino acid when given the native backbone. State-of-the-art models achieve:

• ~52-55% recovery on single-chain proteins (ProteinMPNN)
• ~40-45% recovery on protein-protein interfaces
• ~35-40% recovery on antibody CDR loops Importantly, perfect recovery is not the goal; the model should identify alternative sequences that also fold correctly. Perplexity on held-out test sequences and experimental validation through recombinant expression and biophysical characterization provide complementary evaluation.

Inverse folding models are increasingly paired with backbone generation models (e.g., RFdiffusion, FrameBuilder) in a two-stage design pipeline:

1. A diffusion or hallucination model generates novel backbone structures with desired functional features (binding pockets, symmetry, topology)
2. An inverse folding model designs sequences that encode these structures This modular architecture separates the geometric problem (what shape?) from the sequence problem (what amino acids?), enabling plug-and-play protein design where different backbone generators and sequence designers can be combined for specific engineering goals.

A message-passing neural network designed specifically for the inverse folding problem. ProteinMPNN predicts amino acid sequences that will fold into a given backbone structure with high experimental success rates.

• Uses an encoder-decoder architecture with protein backbone features as input
• Achieves 52% native sequence recovery on average, significantly outperforming physics-based methods
• Robust to structural noise, making it suitable for de novo designed scaffolds
• Outputs position-specific amino acid probabilities for downstream optimization

The computational task of determining the optimal discrete rotameric state for each amino acid side chain on a fixed backbone scaffold. This is a critical post-processing step after inverse folding.

• Searches the rotamer library of statistically preferred side-chain conformations
• Minimizes steric clashes and maximizes favorable van der Waals interactions
• Often coupled with energy minimization to resolve local geometric strain
• Essential for validating that designed sequences are physically realizable

The change in thermodynamic stability of a protein upon mutation, calculated as the difference in Gibbs free energy of folding between mutant and wild-type sequences.

• Negative ΔΔG indicates stabilizing mutations that favor the folded state
• Positive ΔΔG indicates destabilizing mutations that may prevent proper folding
• Used to computationally validate inverse folding predictions before experimental testing
• Tools like FoldX and Rosetta ddg_monomer provide rapid in silico ΔΔG estimation

A collection of structurally distinct states representing the intrinsic dynamic flexibility of a protein. Inverse folding must account for this flexibility rather than targeting a single static structure.

• Proteins exist as ensembles of conformers, not rigid bodies
• Designing for a single snapshot may produce sequences that fail on alternative conformations
• Multi-state design algorithms optimize sequences against multiple backbone states simultaneously
• Critical for designing allosteric proteins and flexible binding interfaces

A high-throughput experimental technique that measures the functional impact of thousands of protein sequence variants simultaneously, generating rich datasets for training and validating inverse folding models.

• Couples massively parallel sequencing with functional selection assays
• Produces comprehensive genotype-phenotype maps for entire protein domains
• Provides ground-truth data on which mutations are tolerated at each position
• Used to benchmark inverse folding algorithms against experimental fitness landscapes

A neural network architecture that guarantees its predictions transform predictably under 3D rotations and translations of input coordinates. This property is fundamental to physically consistent inverse folding.

• Ensures that rotating the backbone produces identically rotated sequence predictions
• SE(3) equivariance preserves the geometric relationships between residues
• Contrasts with invariant networks that discard orientation information
• Core architectural principle in models like ProteinMPNN and ESM-IF1