Structure-Guided
Protein Design

Protein function is determined by three-dimensional structure, not sequence alone. Two proteins with 30% sequence identity can adopt nearly identical folds, while a single point mutation at a critical position can abolish function entirely. Structure-guided design leverages this reality by using 3D models to identify which positions matter, which surfaces are accessible, and which interactions drive binding or stability—before making any sequence changes.

The revolution in structure prediction—AlphaFold2, ESMFold, and more recently Boltz-2—has made high-quality structural models available for virtually any protein sequence within minutes. This means that structure-guided design is no longer limited to proteins with experimental crystal structures. A computationally designed nanobody can be modeled in complex with its target, its binding interface analyzed for shape complementarity and electrostatic compatibility, and its framework assessed for potential stability liabilities, all computationally and in under an hour.

The binding interface between a protein and its target is typically composed of 15–30 residues on each side, but the energetic contribution is unevenly distributed. A small number of “hotspot” residues contribute the majority of binding free energy, while peripheral residues may contribute little or even be slightly unfavorable. Identifying these hotspots—through computational alanine scanning, Rosetta interface energy decomposition, or experimental mutagenesis data—focuses engineering effort on the positions that actually matter.

Computational docking predicts how two proteins interact at the atomic level, revealing the geometry of the binding interface, key hydrogen bonds, salt bridges, and hydrophobic contacts. For antibody-antigen interactions, co-folding tools like Boltz-2 and AlphaFold-Multimer predict the complex structure directly from sequence, providing interface metrics (ipTM, PAE) that correlate with binding confidence. These predictions guide decisions about which epitope to target, which CDR positions to diversify, and which framework mutations might improve interface packing.

The most powerful application of structure-guided design is de novo backbone generation, where AI models like RFDiffusion and BoltzGen create entirely new protein backbones conditioned on a target structure. Rather than searching existing antibody repertoires for binders, these tools generate novel protein folds designed to complement a specific binding surface. The generated backbones are then threaded with sequences using inverse folding models (ProteinMPNN), producing candidates that are structurally plausible and sequence-optimized for folding.

This approach is particularly valuable for targets that have resisted traditional antibody discovery—small proteins with limited surface area, targets with conformational heterogeneity, or epitopes buried in protein-protein interfaces. Structure-guided de novo design can generate hundreds of unique backbone topologies targeting the same epitope, providing a level of design diversity that no experimental library can match.