Generative Protein and Molecule Design

AlphaFold2 answered the question biology had been asking for fifty years: given a sequence, what shape does it fold into (see AlphaFold2 protein structure)? Design asks the inverse and harder question. Given a shape you want — a binder that grips a specific patch of a viral protein, a scaffold that holds three catalytic residues in exact geometry, a symmetric cage that self-assembles — produce a protein that adopts it. There is no natural sequence to look up, because the target does not exist in nature.

Denoising a backbone into existence

RFdiffusion's move is to take a structure prediction network and fine-tune it on a denoising task, turning a discriminative model into a generative one (Watson et al., 2023, De novo design of protein structure and function with RFdiffusion, Nature 620, 1089–1100). Training corrupts real protein backbones with noise over many steps; the network learns to reverse one step of that corruption. At sampling time you start from pure noise and run the reverse process, and a plausible backbone condenses out of it (see diffusion models).

The reason this works well is that the network already knows what proteins look like — it was pretrained to predict their structures — so denoising inherits a strong prior rather than learning geometry from scratch. Conditioning is what makes it useful: the reverse process can be constrained to keep a fixed motif in place, to build around a target's binding surface, or to obey a symmetry group, so the model generates the rest of the protein around requirements you impose.

Structure, then sequence, then verification

RFdiffusion produces backbone coordinates, not amino acids. A second network, ProteinMPNN, performs inverse folding: given a backbone, choose a sequence likely to fold into it. It recovers 52.4% of native sequence on natural backbones against 32.9% for Rosetta's physics-based design (Dauparas et al., 2022, Science 378, 49–56), and its practical contribution was rescuing designs that had previously failed experimentally.

The third step is in silico verification and it defines the working pipeline: fold the designed sequence with a structure predictor and check that the prediction matches the intended backbone. Designs whose predicted structure disagrees with their design target are discarded before anyone touches a pipette. This self-consistency filter is what turns a generative model with a low hit rate into a usable funnel — you generate thousands, filter to dozens, and test those.

The same shape in small molecules

Molecular generative models follow the analogous arc — generate candidate structures, then filter hard — but with a sharper failure mode. It is trivial to produce molecules that score well on a computational objective and are unmakeable, unstable, or exploit a flaw in the scoring function. The field's corrective is to constrain generation to synthesisable chemistry and to score with retrosynthesis feasibility alongside predicted affinity, because a compound no chemist can make has no value regardless of its predicted binding.

When it breaks

• The metric that matters is experimental, and it is low. Design success is measured as the fraction of ordered, expressed, purified candidates that actually bind or fold as intended, and single-digit percentages are normal even for celebrated results. Any paper reporting only computational scores is reporting the easy half.
• Self-consistency filtering is circular in a specific way. The same learned prior generates and evaluates, so designs that exploit a shared blind spot pass the filter and fail in the lab. Success rates measured after filtering are not a probability that the next design works.
• Designed proteins are biased toward what is easy to design. Idealised, hyper-stable, alpha-helical structures are overrepresented, because that is where both the generator and the verifier are most confident. Loops, disorder, and conformational change are underserved.
• Function is not structure. Enzymes need transition-state geometry maintained under motion, not just a static pocket in the right shape; binders need specificity against a whole proteome, not affinity for one target. Getting the backbone right is necessary and a long way from sufficient.
• Dual-use is a live constraint, not a footnote. Design tools that produce novel binders for arbitrary targets carry biosecurity implications that the field is actively negotiating, and access to the strongest models is deliberately governed.