Abstract
The protein structure prediction problem, largely solved for single-chain natural proteins by AlphaFold2, has given way to a harder challenge: designing novel proteins with desired functions from scratch. Diffusion models have emerged as the leading approach for generative protein design, operating directly in the space of backbone coordinates or atomic positions rather than sequence space. RFdiffusion from Baker Lab at the University of Washington, ProteinMPNN as a sequence design complement, and more recent all-atom diffusion approaches have demonstrated that functional proteins, including binders, enzymes, and symmetric assemblies, can be computationally designed with measurable wet-lab success rates.
How Diffusion Models Apply to Protein Geometry
Standard image diffusion corrupts pixel values with Gaussian noise and trains a network to denoise. For proteins, the equivalent forward process corrupts backbone torsion angles or Cartesian coordinates. RFdiffusion operates on backbone frames using SE(3)-equivariant diffusion, meaning the noise and denoising processes respect the rotational and translational symmetries of three-dimensional space. This is important because a protein’s function is determined by its three-dimensional shape, and a generative model that does not respect geometric symmetry would produce physically inconsistent outputs. The RoseTTAFold2 network architecture underlying RFdiffusion handles equivariance through invariant point attention borrowed from AlphaFold2’s structure module.
Conditional Generation and Functional Targeting
Unconditional backbone generation produces geometrically plausible proteins, but practical design requires conditioning: generate a binder that docks to a specific target epitope, or an enzyme with an active site matching a desired catalytic geometry. RFdiffusion supports partial diffusion, where some regions (such as a target protein structure) are held fixed while the diffusion process generates complementary binding partners around them. This motif scaffolding mode has been used to design binders to several medically relevant targets, including influenza hemagglutinin and nanobody-sized scaffolds that retain binding affinity in experimental validation. Sequence design following structure generation is handled by a separate model, typically ProteinMPNN, which conditions amino acid selection on the designed backbone geometry.
Current Limits and Directions
Success rates for computationally designed proteins, measured as the fraction of designs that fold as predicted and perform the intended function, vary widely by task. Binder design against well-characterized targets has reached rates of 10-30% in favorable cases. Enzyme design with novel catalytic activity remains harder, with success rates typically below a few percent before experimental optimization rounds. All-atom diffusion models that jointly generate backbone and side-chain positions, such as work from Generate Biomedicines and Profluent, are an active direction for improving designability. Integration with AlphaFold3’s unified structure prediction, which handles protein-ligand and protein-nucleic acid complexes, may enable conditional generation for a broader class of functional targets over the next year.