AI protein design

AI protein design covers two distinct uses of machine learning on proteins: predicting the three-dimensional structure a given amino acid sequence folds into, and generating new sequences that fold into a shape chosen in advance. The first was largely solved for single chains between 2018 and 2021, mostly by DeepMind's AlphaFold. The second is being solved unevenly, by generative models from David Baker's laboratory and others, and it is the half that produces molecules which have never existed. AlphaFold predicts; it does not design.

§Two problems, not one

Prediction rests on Christian Anfinsen's finding that a small protein's sequence contains the information needed to specify its folded structure. Reading that information out resisted fifty years of physics-based attack. From 1994 the field measured its progress through CASP, a biennial blind assessment run by John Moult in which groups predict structures already solved experimentally but not yet published. AlphaFold2's entry at CASP14 in 2020 returned a median score above 90 on the assessment's main accuracy metric, close to the level at which two experimental determinations of the same protein differ from each other.¹

Design is the inverse problem, with no Anfinsen result to guarantee it is solvable. The Baker laboratory's Top7, reported in 2003, was the first protein designed to a fold not found in nature whose crystal structure matched the design.² For the next fifteen years design meant energy-function optimisation in Rosetta, at success rates low enough that each published design stood on much failed protein.

§How the systems work

AlphaFold2 takes a sequence, retrieves a multiple sequence alignment of its evolutionary relatives, and reasons jointly over that alignment and over a matrix of residue pairs before predicting atomic coordinates directly. The alignment does much of the work: residues that mutate in a correlated way across species tend to touch in three dimensions. AlphaFold3, published in 2024, replaced the final geometric module with a diffusion-based generator and extended the model to complexes of proteins with DNA, RNA, small molecules and ions.³ Protein language models trained on sequence alone drop the alignment, trading some accuracy for speed and for coverage of sequences with no known relatives.

Design runs in two stages. A generative model such as RFdiffusion, adapted from a structure predictor by training it to denoise corrupted backbones, produces a protein shape to order: a binder against a specified epitope, a symmetric assembly, or a scaffold holding catalytic residues in position.⁴ A second model, ProteinMPNN, then reads that backbone and proposes sequences likely to fold into it, raising experimental success rates well above what Rosetta's energy functions achieved.⁵ Designs are filtered in silico by running a structure predictor on the proposed sequence and asking whether it folds back to the intended shape; survivors are ordered as synthetic genes, expressed, and tested. That filter is partly circular, since a design that fools the predictor is selected for alongside one that works.

What "de novo" means here

A de novo designed protein is one whose sequence was chosen computationally rather than copied from a natural one. It is not a synonym for "new drug": most engineered biologics, including nearly all therapeutic antibodies, are natural scaffolds modified at a few positions.

§Development history

1. 1972
   Sequence determines structure
   Christian Anfinsen shares the Nobel Prize in Chemistry for work showing that a denatured enzyme refolds spontaneously, implying the fold is encoded in the sequence.
2. 1994
   CASP begins
   John Moult launches a biennial blind assessment in which groups predict structures already solved but not yet released, giving the field an honest scoreboard.
3. 2003
   Top7
   Kuhlman and colleagues in the Baker laboratory design a protein with a fold not seen in nature and confirm it crystallographically.
4. 2018
   AlphaFold at CASP13
   DeepMind's first entry places first among predictors while remaining well short of experimental accuracy.
5. 2020
   CASP14
   AlphaFold2 predicts most single-chain targets at near-experimental accuracy, and the assessors describe the problem as largely solved for such targets.
6. 2021
   Methods published and opened
   AlphaFold2 is published with its code released; the Baker laboratory publishes RoseTTAFold, a smaller three-track network, within weeks.
7. 2022
   Prediction becomes infrastructure
   The AlphaFold Protein Structure Database expands to over 200 million predicted structures, and ProteinMPNN makes sequence design routine.
8. 2023
   Generative backbones
   RFdiffusion generates protein backbones to order, including binders against chosen targets and scaffolds for catalytic sites.
9. 2024
   AlphaFold3 and the Nobel
   AlphaFold3 extends prediction to biomolecular complexes; the Nobel Prize in Chemistry goes half to David Baker for computational protein design and half jointly to Demis Hassabis and John Jumper for structure prediction.

The 2024 prize is the clearest official statement that these are two achievements rather than one. Half went to David Baker for computational protein design; the other half was shared by Demis Hassabis and John Jumper for structure prediction.⁶

§Current state

Prediction has become laboratory infrastructure. The AlphaFold database offers a predicted structure for essentially every sequence in the major reference databases, and models are used routinely to generate hypotheses about uncharacterised proteins and to help solve experimental structures. That the code and the database are free is unusual for a result of this weight, and helps explain why it diffused in months rather than years.

Design has produced working molecules in narrower domains. Miniprotein binders against viral surface proteins were designed and shown to bind and neutralise in laboratory assays within months of a target sequence becoming available. Computationally designed self-assembling nanoparticles, of a kind that also interests workers in DNA nanotechnology, serve as vaccine scaffolds: a two-component designed particle displaying a coronavirus receptor-binding domain became a vaccine approved in South Korea in 2022, the clearest case of a designed protein reaching people inside an approved product.⁷ Designed enzymes exist, including luciferases built around a chosen chemistry, though catalytic efficiencies for de novo designed enzymes generally remain well below those of their natural counterparts.

The methods are also being turned on the proteins other biotechnologies depend on: variants of the nucleases behind CRISPR–Cas9 and Base editing, generated by sequence models rather than found in bacteria; engineered capsids that redirect AAV vectors used in Somatic gene therapy; and designed binding domains for Epigenome editing. NewLimit applies related machine learning to a different question, which transcription factor combinations to deliver.

As of 2026 no de novo designed protein has been approved anywhere as a therapeutic drug. The South Korean vaccine is the nearest case, and even there the designed part is the scaffold rather than the antigen the immune system is being trained on. A designed mimic of interleukin-2, reported in 2019 and meant to separate the cytokine's antitumour signalling from its toxicity, reached early-phase clinical testing in patients before its developer discontinued the programme.⁸

§Limitations

A predicted structure is not a predicted function. The models return coordinates, not a mechanism, a binding partner, or a catalytic rate, and the shape of an uncharacterised protein often leaves its biology as obscure as before. They also return one static conformation, when what matters is frequently the movement between conformations or the response to a ligand.

Three failure modes are well documented. Intrinsically disordered regions, a large fraction of the human proteome, have no single structure to predict; the models flag them with low confidence, which is useful, but a confidence score is not a description. Prediction of complexes, including antibody–antigen pairs, remains substantially worse than prediction of single chains, which is one reason CASP has kept running. Most importantly for medicine, the models are largely insensitive to single amino acid substitutions: a sequence carrying a destabilising disease mutation usually returns almost the same confident structure as the healthy one.⁹ Benchmark studies of drug docking have likewise found predicted structures to perform worse than experimental ones.

Structure is not evidence

Confidence scores estimate how closely a model matches what an experiment would find, not whether the protein does what a paper claims. Predicted structures have already entered the literature as though they were determinations, and a confidently wrong model is harder to catch than an obviously wrong one.

Design has its own ceiling. Success rates for de novo binders against a new target remain low and vary sharply from target to target, so the method depends on high-throughput screening rather than replacing it. Function beyond binding is much harder than shape. And because designed sequences do not occur in nature, their behaviour in a human immune system is hard to predict in advance: anti-drug antibodies are a familiar problem even for biologics built from human proteins.

§Risks

Reversibility depends on delivery. A designed protein given as a drug clears from the body like any other biologic; the same protein expressed from a delivered gene, or built into an organism that replicates, does not.

The biosecurity concern is specific. Orders for synthetic DNA are screened by many providers against databases of sequences of concern, and that screening is one of the few practical chokepoints in Dual-use research of concern on pathogens and toxins. A 2025 study led by researchers at Microsoft showed that open design tools could generate variants of toxin proteins that evaded existing screening software; patches were distributed to screening providers before publication.¹⁰ It is read both as evidence that the risk is real and that the field can respond, and it sits alongside questions about Synthetic genomes, Genetic code expansion and recoding and Mirror life on the agenda of biosecurity bodies.

Researchers in the field signed voluntary commitments in 2024 to synthesis screening and to restraint in releasing certain capabilities. The comparison usually drawn is with the Asilomar Conference on Recombinant DNA of 1975, and it carries the same objection: self-governance worked when the community controlled the reagents, and open model weights are not controlled that way. Catastrophic-risk analysts generally rank engineered pathogens above designed proteins as a near-term concern, though the tools overlap.

§Outlook

Two questions decide how much the technology matters. The first is whether prediction can be pushed from static structures to the quantities that determine whether a molecule works: binding affinity, conformational ensembles, and the effect of a mutation. The second is whether design can deliver function rather than form, since a de novo enzyme matching a natural one would be a stronger result than any binder.

The deflationary case deserves stating. Producing candidate molecules was never the slowest step in developing a drug; recruitment and clinical failure are, and the largest single cause of that failure is lack of efficacy rather than any property of the molecule, which better structures do not address. The same argument shapes debate over AI drug discovery and over whether machine-designed delivery vehicles and nanoparticle formulations shorten timelines in practice. Demis Hassabis has presented AlphaFold as an early demonstration of what progress toward Artificial general intelligence would buy for science. The counter-argument is not that the achievement is small but that the problem was unusually well posed: a discrete input, a checkable output, and fifty years of curated experimental structures to learn from. Most of biology offers none of those.

§See also

• AI drug discovery
• CRISPR–Cas9
• Synthetic genomes
• Dual-use research of concern
• Mirror life
• AAV vectors
• DNA nanotechnology
• Artificial general intelligence

§References