Loading
What changed between revision 2 and revision 4, compared over the article's markdown source with its frontmatter — so a change to status, evidence or reversibility shows up as the edit it is.
Older · revision 2
Import
Re-import of content/ai-protein-design.md — the file changed after the initial import. Not an edit.
17.2 kB
Revision 2
Revision 4
lines 8–14 → 8–147 unchanged lines not shown
categories: ["genetics", "foundations"]categories: ["genetics", "foundations"]tags: ["machine learning", "protein design", "structural biology", "alphafold", "biosecurity", "drug discovery"]tags: ["machine learning", "protein design", "structural biology", "alphafold", "biosecurity", "drug discovery"]summary: "Machine-learning systems that predict what shape a protein sequence folds into and that generate new proteins with no natural counterpart."summary: "Machine-learning systems that predict what shape a protein sequence folds into and that generate new proteins with no natural counterpart."updated: "2026-07-27"updated: "2026-10-29"humanEvidence: "A computationally designed nanoparticle scaffold is a component of a COVID-19 vaccine approved for use in people in South Korea, and a designed cytokine mimic reached early-phase human trials; no de novo designed protein is approved as a therapeutic."humanEvidence: "A computationally designed nanoparticle scaffold is a component of a COVID-19 vaccine approved for use in people in South Korea, and a designed cytokine mimic reached early-phase human trials; no de novo designed protein is approved as a therapeutic."access: "Most prediction and design code is open source and the AlphaFold structure database is public, so the models are free to run; the gene synthesis and wet-lab testing that turn a design into a molecule are not."access: "Most prediction and design code is open source and the AlphaFold structure database is public, so the models are free to run; the gene synthesis and wet-lab testing that turn a design into a molecule are not."reversibility: "context"reversibility: "context"lines 72–78 → 72–7957 unchanged lines not shown
AlphaFold2 takes a sequence, retrieves a multiple sequence alignment of its evolutionary relatives,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 atomicand reasons jointly over that alignment and over a matrix of residue pairs before predicting atomiccoordinates directly. The alignment does much of the work: residues that mutate in a correlated waycoordinates directly. The alignment does much of the work: residues that mutate in a correlated wayacross species tend to touch in three dimensions. AlphaFold3, published in 2024, replaced the finalacross species tend to touch in three dimensions. AlphaFold3, published in 2024 by DeepMind and itsdrug-discovery sister company [[isomorphic-labs|Isomorphic Labs]], replaced the finalgeometric module with a diffusion-based generator and extended the model to complexes of proteinsgeometric module with a diffusion-based generator and extended the model to complexes of proteinswith DNA, RNA, small molecules and ions.[^abramson2024] Protein language models trained on sequencewith DNA, RNA, small molecules and ions.[^abramson2024] Protein language models trained on sequencealone drop the alignment, trading some accuracy for speed and for coverage of sequences with noalone drop the alignment, trading some accuracy for speed and for coverage of sequences with nolines 119–125 → 120–12740 unchanged lines not shown
for essentially every sequence in the major reference databases, and models are used routinely tofor essentially every sequence in the major reference databases, and models are used routinely togenerate hypotheses about uncharacterised proteins and to help solve experimental structures. Thatgenerate hypotheses about uncharacterised proteins and to help solve experimental structures. Thatthe code and the database are free is unusual for a result of this weight, and helps explain why itthe code and the database are free is unusual for a result of this weight, and helps explain why itdiffused in months rather than years.diffused in months rather than years. AlphaFold3's code, by contrast, followed publication by sixmonths and only for non-commercial use, while its co-developer uses the model commercially. Design has produced working molecules in narrower domains. Miniprotein binders against viral surfaceDesign has produced working molecules in narrower domains. Miniprotein binders against viral surfaceproteins were designed and shown to bind and neutralise in laboratory assays within months of aproteins were designed and shown to bind and neutralise in laboratory assays within months of alines 153–159 → 155–16127 unchanged lines not shown
Three failure modes are well documented. Intrinsically disordered regions, a large fraction of theThree failure modes are well documented. Intrinsically disordered regions, a large fraction of thehuman proteome, have no single structure to predict; the models flag them with low confidence, whichhuman proteome, have no single structure to predict; the models flag them with low confidence, whichis useful, but a confidence score is not a description. Prediction of complexes, includingis useful, but a confidence score is not a description. Prediction of complexes, includingantibody–antigen pairs, remains substantially worse than prediction of single chains, which is oneantibody-antigen pairs, remains substantially worse than prediction of single chains, which is onereason CASP has kept running. Most importantly for medicine, the models are largely insensitive to singlereason CASP has kept running. Most importantly for medicine, the models are largely insensitive to singleamino acid substitutions: a sequence carrying a destabilising disease mutation usually returnsamino acid substitutions: a sequence carrying a destabilising disease mutation usually returnsalmost the same confident structure as the healthy one.[^buel2022] Benchmark studies of drug dockingalmost the same confident structure as the healthy one.[^buel2022] Benchmark studies of drug dockinglines 197–216 → 199–22037 unchanged lines not shown
Two questions decide how much the technology matters. The first is whether prediction can be pushedTwo questions decide how much the technology matters. The first is whether prediction can be pushedfrom static structures to the quantities that determine whether a molecule works: binding affinity,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 deliverconformational ensembles, and the effect of a mutation. Isomorphic Labs claimed such gains forbinding affinity in 2026, in a company technical report rather than a peer-reviewed paper. The second is whether design can deliverfunction rather than form, since a de novo enzyme matching a natural one would be a stronger resultfunction rather than form, since a de novo enzyme matching a natural one would be a stronger resultthan any binder.than any binder. The deflationary case deserves stating. Producing candidate molecules was never the slowest step inThe deflationary case starts further down the pipeline. Producing candidate molecules was never thedeveloping a drug; recruitment and clinical failure are, and the largest single cause of thatslowest step in developing a drug; recruitment and clinical failure are, and the largest singlefailure is lack of efficacy rather than any property of the molecule, which better structures docause of that failure is lack of efficacy rather than any property of the molecule, which betternot address. The same argument shapes debate over [[ai-drug-discovery]] and over whether machine-designedstructures do not address. The same argument shapes debate over [[ai-drug-discovery]] and over[[targeted-drug-delivery|delivery vehicles]] and [[lipid-nanoparticles|nanoparticle formulations]]whether machine-designed [[targeted-drug-delivery|delivery vehicles]] andshorten timelines in practice. Demis Hassabis has presented AlphaFold as an early demonstration of[[lipid-nanoparticles|nanoparticle formulations]] shorten timelines in practice. Demis Hassabis haswhat progress toward [[artificial-general-intelligence]] would buy for science. The counter-argumentpresented AlphaFold as an early demonstration of what progress towardis not that the achievement is small but that the problem was unusually well posed: a discrete input,[[artificial-general-intelligence]] would buy for science. The counter-argument grants thea checkable output, and fifty years of curated experimental structures to learn from. Most ofachievement and holds that the problem was unusually well posed: a discrete input, a checkablebiology offers none of those.output, and fifty years of curated experimental structures to learn from. Most of biology offersnone of those. ## See also## See also removed, struck through added, underlinedLine numbers count the serialised markdown of each revision, frontmatter included.
223 lines are the same in both revisions, and 190 of them are not shown. Neither revision here is necessarily the current article — read AI protein design for the text this wiki stands behind today.