Inverse Folding

The core mechanism of inverse folding models is generating an amino acid sequence conditioned on a fixed 3D backbone structure. Unlike forward folding (AlphaFold), which predicts structure from sequence, these models take backbone coordinates as input and output a probability distribution over the 20 canonical amino acids at each position. The model must learn the complex relationship between local and non-local structural environments and the physicochemical properties of residues that stabilize them.

• Input: 3D coordinates of N, Cα, C, and optionally O atoms
• Output: A probability vector over 20 amino acids per residue position
• Key challenge: Capturing both local backbone geometry and long-range spatial contacts

Leading inverse folding models like ProteinMPNN represent the input backbone as a k-nearest neighbor graph where nodes are residues and edges connect spatially proximal Cα atoms. Message-passing neural networks then iteratively update residue representations by aggregating information from neighboring nodes. This architecture naturally captures the spatial relationships critical for determining which amino acids can pack together without steric clashes.

• Graph construction: Edges defined by Cα-Cα distance, typically k=30-48 nearest neighbors
• Edge features: Pairwise distances and relative orientation vectors between residue frames
• Message passing: Multiple layers of structured updates incorporating both node and edge features
• Key advantage: Inherently invariant to global rotation and translation of the input structure

Inverse folding models employ different strategies for generating the full sequence. Autoregressive decoding predicts residues sequentially, conditioning each prediction on previously generated amino acids, which captures epistatic interactions between sequence positions. One-shot prediction generates all positions independently in parallel, offering faster inference but potentially missing cooperative effects.

• ProteinMPNN: Autoregressive with random decoding order, which acts as a form of data augmentation during training
• ESM-IF1: One-shot prediction using invariant geometric processing
• Trade-off: Autoregressive models better capture sequence dependencies; one-shot models are computationally faster
• Temperature sampling: Both approaches allow tuning of sequence diversity via softmax temperature

The primary evaluation metric for inverse folding models is native sequence recovery rate—the percentage of residues where the model's predicted amino acid matches the wild-type sequence for a given backbone. High recovery indicates the model has learned the sequence-structure relationship encoded in natural evolution. However, perfect recovery is not always desirable, as multiple sequences can stably fold into the same backbone.

• Typical performance: ProteinMPNN achieves ~52% recovery on single-chain proteins
• Perplexity: Also used to measure the model's uncertainty over the sequence distribution
• Limitation: Recovery rate does not measure whether designed sequences actually fold correctly
• Experimental validation: Designs must be tested via recombinant expression and biophysical characterization

A critical practical feature of modern inverse folding models is robustness to structural noise. Real protein backbones—whether from crystal structures, NMR ensembles, or computational design—contain coordinate errors. ProteinMPNN was explicitly trained with Gaussian noise added to backbone coordinates, making it tolerant of imprecise input structures. This noise augmentation prevents the model from overfitting to exact atomic positions and improves generalization to novel backbone geometries.

• Training augmentation: 0.1–0.3 Å Gaussian noise on Cα coordinates
• Benefit: Enables design directly from low-resolution cryo-EM density maps
• Backbone flexibility: Some models accept ensembles of conformations rather than single static structures
• Practical impact: Reduces failure rate when designing for computationally generated backbones with minor geometric imperfections

Advanced inverse folding models extend beyond single-chain proteins to design sequences for multi-chain complexes and symmetric assemblies. ProteinMPNN can simultaneously design sequences for all chains in a complex, accounting for inter-chain interfaces. For symmetric systems like viral capsids or designed nanocages, symmetry constraints can be enforced to ensure identical sequences at symmetry-related positions.

• Multi-chain input: Separate chain identifiers with inter-chain edges in the k-NN graph
• Interface optimization: Model learns to place hydrophobic residues at buried interfaces and hydrophilic residues at solvent-exposed surfaces
• Symmetry handling: Cyclic, dihedral, and icosahedral symmetries supported via tied sequence constraints
• Application: Design of self-assembling protein nanomaterials and logic-gated therapeutic complexes

The computational creation of entirely new protein sequences and structures that do not exist in nature, designed to perform a specific function. Inverse folding is a critical sub-task here, as it bridges the gap between a desired novel backbone and a viable sequence. This field combines Rosetta-based design with deep learning to create proteins for therapeutics, catalysis, and self-assembling nanomaterials.

The computational process of predicting the optimal 3D conformations of amino acid side chains onto a fixed protein backbone, often using discrete rotamer libraries. While inverse folding selects the amino acid identity, side-chain packing resolves the specific chi angles to minimize steric clashes and maximize favorable interactions like hydrogen bonds and van der Waals contacts.

A 2D plot of the phi (φ) and psi (ψ) backbone dihedral angles of amino acid residues in a protein structure, used to validate the stereochemical quality of a model. For inverse folding, the target backbone must have residues falling within allowed regions. Generated sequences are often evaluated by how well their predicted structures recapitulate favorable Ramachandran distributions.

A deep learning system developed by DeepMind that predicts a protein's 3D structure from its amino acid sequence with atomic accuracy. It serves as the forward folding counterpart to inverse folding. In design pipelines, AlphaFold is often used to validate whether the sequence generated by an inverse folding model actually folds into the intended target backbone.

A high-throughput experimental technique that measures the functional effect of thousands of genetic variants, providing massive datasets for training variant effect predictors. DMS data provides a ground-truth benchmark for inverse folding models, testing whether they can predict which mutations are tolerated at a given position without disrupting the fold.