Protein design is the construction of new protein sequences that fold into an intended three-dimensional shape and perform an intended job, such as binding a chosen target, catalyzing a reaction, or assembling into a larger structure. It is the inverse of the protein folding problem: instead of asking what shape a given sequence adopts, a designer asks which sequence will adopt a chosen shape. Early work relied on physics-based energy functions and Monte Carlo search, most prominently in the Rosetta software suite developed in David Baker's laboratory at the University of Washington. Since 2020, deep learning has largely displaced those methods. Current pipelines combine generative models of protein structure, inverse-folding networks that write sequences onto a generated backbone, and structure predictors used as in silico filters.

The field's public standing rose sharply on October 9, 2024, when the Royal Swedish Academy of Sciences awarded half of the Nobel Prize in Chemistry to Baker "for computational protein design" and the other half jointly to Demis Hassabis and John Jumper of Google DeepMind "for protein structure prediction." [1] The pairing was apt: AlphaFold made structure prediction cheap enough to use as a design filter, which is one reason design became practical in roughly the same window.

From Rosetta to Top7

Rosetta began in Baker's lab in the late 1990s as a fragment-assembly method for ab initio structure prediction, and the name first appeared in the literature around the CASP3 assessment in 1998. [2] Because the energy function that scores a predicted fold can also score a candidate sequence on a fixed backbone, the prediction code grew a design branch. In 2005 the lab launched Rosetta@home, a volunteer computing project on the BOINC platform that donated large amounts of distributed compute to the problem, and in 2008 it released Foldit, a puzzle game in which players manipulate protein structures by hand. Foldit players went on to contribute to published structure and enzyme-improvement results. [3][4]

The landmark early design result was Top7, reported in Science in 2003 by Brian Kuhlman, Gautam Dantas and colleagues. Top7 is a 93-residue alpha/beta protein with a sequence and a topology not found in the Protein Data Bank. It expressed, folded, proved unusually stable, and its crystal structure matched the design model to a root-mean-square deviation of 1.2 angstroms. [5] Top7 showed that a fold could be invented rather than borrowed from nature. It had no function.

Function arrived slowly. In 2008 Daniela Rothlisberger, Olga Khersonsky and co-workers reported computationally designed Kemp eliminases in Nature, enzymes that catalyze a reaction with no natural counterpart. Eight designs were active, with rate enhancements up to roughly 10^5 over the uncatalyzed reaction, and laboratory evolution then improved catalytic efficiency more than 200-fold. [6] The designs worked, but they were far weaker than natural enzymes, and that gap persisted for more than a decade.

Two later results showed what design could offer medicine. In 2019 Daniel-Adriano Silva and colleagues described Neo-2/15, a de novo protein that reproduces the receptor-binding surface of the cytokines interleukin-2 and interleukin-15 while sharing almost no sequence or topology with either. Neo-2/15 binds the IL-2Rbeta/gamma-c receptor more tightly than the natural cytokines, does not engage the alpha subunits associated with IL-2 toxicity, and outperformed IL-2 in mouse melanoma and colon cancer models with lower toxicity. [7] Separately, the Institute for Protein Design built self-assembling protein nanoparticles that display viral antigens. One of them, developed with SK bioscience and GSK as the COVID-19 vaccine GBP510, was approved by South Korea's Ministry of Food and Drug Safety in June 2022 under the brand name SKYCovione, cleared in the United Kingdom in May 2023, and granted a World Health Organization Emergency Use Listing in June 2023. The Institute for Protein Design calls it the first computationally designed protein medicine to win approval, a characterization that comes from the developers rather than from an independent adjudicator. [8][9]

How a modern design pipeline works

Most current pipelines split the problem into three computational stages and then test the survivors in a laboratory.

Backbone generation produces a three-dimensional shape without yet deciding which amino acid sits at each position. Diffusion-based generators dominate this step: they start from noise and iteratively denoise into a protein-like geometry, optionally conditioned on a binding site, a symmetry requirement, a fixed functional motif, or an overall shape.

Sequence design, also called inverse folding, assigns amino acids to that fixed backbone. The standard tool is ProteinMPNN, a graph neural network that treats residues as nodes and predicts an identity for each position. Reported in Science in 2022 by Justas Dauparas and colleagues, it raised the fraction of designs that fold as intended well above what Rosetta's physics-based sequence design achieved, and it runs in about a second per design on a single GPU. [10]

In silico filtering discards most candidates before anything is synthesized. The usual test is to feed the designed sequence back into a structure predictor and ask whether the prediction recovers the intended backbone with high confidence. Adding AlphaFold-based filters in this way improved binder design success rates roughly tenfold over unfiltered Rosetta pipelines. [11] Because synthesis and assay capacity is the real bottleneck, the quality of this filter largely determines how much laboratory work a project costs.

Experimental validation still settles the question. Designs are expressed in bacteria, yeast or mammalian cells, screened for binding by yeast display or surface plasmon resonance, measured for affinity, and in the best cases confirmed by X-ray crystallography or cryo-electron microscopy against the design model.

The deep learning era

AlphaFold 2 reached a median Global Distance Test score of 92.4 at CASP14 in 2020, the first time a method matched experimental accuracy across most targets. [12] RoseTTAFold, published in Science in July 2021 by Minkyung Baek and colleagues in the Baker lab, reproduced much of that accuracy with an open implementation and became the base network for the design tools that followed. [13] The consequence for design was straightforward but not obvious in advance: a fast, accurate predictor is a cheap simulator of the folding step, so a design program can score millions of candidates without touching a pipette.

RFdiffusion, published in Nature in July 2023, adapted the denoising diffusion framework from image generation to protein backbones by fine-tuning RoseTTAFold. It produced monomers, symmetric oligomers, metal-binding proteins, enzyme active-site scaffolds and target-specific binders, and it became the default open tool for de novo design. [14] Chroma, published in Nature in November 2023 by John Ingraham and colleagues at Generate Biomedicines, took a related approach with a correlated diffusion process tuned to polymer statistics, plus conditioning on symmetry, substructure, shape and text-style semantic prompts. The authors characterized 310 designs experimentally and solved two crystal structures that matched the samples at roughly 1.0 angstrom backbone accuracy. [15]

AlphaProteo, announced by Google DeepMind in September 2024, generates binder structure and sequence jointly rather than in two stages. Across seven targets it reported binding affinities 3 to 300 times better than the best prior computational methods, including an 88 percent experimental success rate on one viral target, and it failed outright on TNF-alpha. The machine learning methods were withheld for stated biosecurity and commercial reasons, and the system was never released as software. [16]

A parallel line of work treats proteins as text. Protein language models such as the ESM family train on hundreds of millions of raw sequences with masked-token objectives borrowed from large language model research, and they pick up structural and functional regularities without being told anything about geometry. ESMFold showed that such a model can predict structure without a multiple sequence alignment. ESM3, released by EvolutionaryScale on June 25, 2024 and published in Science in 2025, reasons jointly over sequence, structure and function in a single network, with a flagship version at 98 billion parameters and a smaller 1.4 billion parameter checkpoint released with open weights. Its launch demonstration was esmGFP, a fluorescent protein about 58 percent identical to its closest natural relative, a distance the authors equated to roughly 500 million years of natural evolution. [17][18]

Catalysis, the hardest target, moved in 2025. Anna Lauko, Sam Pellock and colleagues reported computationally designed serine hydrolases in Science, built by pairing RFdiffusion with an evaluation step that checked whether a designed active site held together across the whole reaction coordinate rather than in a single snapshot. The best designs reached catalytic efficiencies around 2 x 10^5 M^-1 s^-1 on five folds unlike those of natural serine hydrolases, with crystal structures within 1 angstrom of the design models. Pellock still described them as "not quite as good as native enzymes." [19]

The pace has not slowed. RFdiffusion2, announced by the Institute for Protein Design in April 2025 and published in Nature Methods in December 2025, scaffolds enzyme active sites at the atom level; on a 41-case enzyme benchmark it scaffolded all 41 sites where earlier methods managed 16. [20] RFdiffusion3, released as open source in December 2025, extends the approach to all-atom design of interactions with DNA and other molecules, though its core paper remains a preprint. [21] Boltz-2, released in June 2025 under an MIT license by researchers at MIT and Recursion Pharmaceuticals, paired structure prediction with a binding-affinity module. [22]

Notable models and systems

System	Group	Released	Role in design
Rosetta	Baker lab, University of Washington	1998 onward	Physics-based folding and sequence design; produced Top7 [2][5]
AlphaFold 2	Google DeepMind	2020-2021	Structure prediction; became the standard design filter [12]
RoseTTAFold	Baker lab	2021	Open structure predictor; backbone network for RFdiffusion [13]
ProteinMPNN	Baker lab	2022	Inverse folding: sequences for a fixed backbone [10]
RFdiffusion	Baker lab	2023	Diffusion generator for backbones and binders [14]
Chroma	Generate Biomedicines	2023	Programmable diffusion generator with rich conditioning [15]
AlphaFold 3	Google DeepMind and Isomorphic Labs	2024	Joint prediction of proteins with DNA, RNA, ligands and ions [23]
ESM3	EvolutionaryScale	2024	Generative model over sequence, structure and function [17][18]
AlphaProteo	Google DeepMind	2024	Joint binder structure and sequence generation [16]
BindCraft	Correia lab, EPFL	2025	Open one-shot binder design pipeline [24]
RFdiffusion2	Institute for Protein Design	2025	Atom-level enzyme active-site scaffolding [20]
Chai-2	Chai Discovery	2025	Zero-shot de novo antibody and miniprotein design [25]

Companies

Protein design drew unusually large private investment between 2021 and 2026, partly because a designed binder is a plausible drug and partly because the academic methods transfer easily into company pipelines.

Company	Origin	Notable funding	Focus
Generate Biomedicines	2018, inside Flagship Pioneering	Around $400 million in a February 2026 Nasdaq initial public offering, after roughly $700 million privately	Generative design of antibodies and other protein therapeutics; built Chroma [15][26]
Xaira Therapeutics	Launched April 23, 2024	More than $1 billion committed at launch	Machine learning drug discovery; David Baker is a co-founder, and the company recruited researchers behind RFdiffusion [27]
EvolutionaryScale	2023, founded by former Meta AI researchers	More than $142 million seed round in June 2024; team joined the Chan Zuckerberg Initiative's Biohub in November 2025	Protein language models; built the ESM3 family [28][29]
Isomorphic Labs	2021, spun out of DeepMind	$600 million in March 2025 and a reported $2.1 billion round in May 2026	Drug design on AlphaFold 3 and the IsoDDE engine [30]
Profluent	2022, Berkeley, California	Venture-backed	Protein language models for genome editors and enzymes; produced OpenCRISPR-1
Chai Discovery	Early 2024, San Francisco	$130 million Series B in December 2025	Structure prediction and antibody design; Chai-1 and Chai-2 [25]
Latent Labs	2023, out of stealth February 2025	$50 million disclosed at launch	Browser-based binder design; the Latent-X models
Nabla Bio	2020, from the Church lab at Harvard	Venture-backed, with a second Takeda collaboration in October 2025	Antibody and multispecific binder design with its JAM models
Cradle	2021, Amsterdam	$73 million Series B in November 2024	Protein engineering software sold to other laboratories
Basecamp Research	2019, London	Venture-backed	Biodiversity sequence data used to train protein models

Clinical status

Very little designed protein has reached patients, and the gap between laboratory success and clinical evidence is the honest summary of the field in 2026.

The clearest case of a fully de novo designed protein entering human trials is NL-201, the clinical version of Neo-2/15, developed by Neoleukin Therapeutics after a spinout from the Institute for Protein Design. Neoleukin stated that it believed NL-201 was the first fully de novo protein to be evaluated in clinical trials. A Phase 1 study in relapsed or refractory solid tumors ran alone and in combination with pembrolizumab, and the company discontinued development in November 2022 for strategic reasons, reporting target engagement and the expected pharmacodynamic changes but no significant immunogenicity. [31] The claim of being first came from the company, not from an outside body.

Most programs now described as AI-designed are antibodies optimized with machine learning rather than new folds. Generate Biomedicines' lead asset GB-0895, a long-acting anti-TSLP antibody, completed a Phase 1 study in 96 patients with mild to moderate asthma and entered two global Phase 3 trials in severe asthma, SOLAIRIA-1 and SOLAIRIA-2, with roughly 1,600 patients planned and the first Phase 3 patient dosed on January 26, 2026. [26] Absci reported interim Phase 1 data for its anti-TL1A antibody ABS-101 in November 2025 and said it would seek a partner rather than continue internal clinical development past Phase 1. [32] Isomorphic Labs, the most heavily funded company in the adjacent structure prediction space, had not dosed a human as of mid-2026; Hassabis said in January 2026 that first trials were expected by the end of that year, a slip from an earlier target of end of 2025. [33]

Benchmarks and evaluation

Protein design has no equivalent of CASP. The CASP assessments through CASP16 in 2024 cover structure prediction, complexes, accuracy estimation, nucleic acids, protein-ligand complexes and conformational ensembles, with no design category. [34] Design results are therefore reported target by target, by whoever made them, which makes comparison awkward.

The common measures are the experimental success rate, meaning the fraction of synthesized designs that bind or catalyze as intended, and the best affinity or catalytic efficiency achieved. Both depend heavily on how many candidates were filtered, which assay was used, and how the target was chosen. BindCraft, published in Nature in 2025 by Martin Pacesa and colleagues at EPFL, is one of the few peer-reviewed one-shot pipelines with a broad report: experimental success rates of 10 to 100 percent depending on target, with nanomolar affinities and no high-throughput screening. [24] Chai Discovery's Chai-2 technical report describes a 16 percent hit rate for fully de novo antibody design across 52 targets with 20 or fewer designs each, a company-reported figure with no independent replication located. [25]

Independent evaluation has started to appear. ProtDBench, an ICML 2026 benchmark, argues that reported in silico metrics are not comparable across methods because evaluation protocols differ, and that the structure predictors used as verifiers are biased and disagree with one another even under identical filters. It benchmarks open generative methods on a fixed set of targets under a fixed compute budget. It does not claim that published success rates are overstated. [35]

Limitations

The input requirement is strict. Binder design generally needs a high-resolution experimental structure of the target, which excludes a large share of medically interesting proteins, and designs against difficult surfaces still fail. AlphaProteo's complete failure on TNF-alpha, a trimeric cytokine targeted by several of the best-selling drugs ever sold, is the cleanest published example. [16]

Enzymes remain much harder than binders. Before RFdiffusion2, methods scaffolded only 16 of 41 active sites in the benchmark its authors used, and the designed serine hydrolases of 2025, the strongest result in the field, are still slower than their natural counterparts. [19][20]

Specificity is under-tested. A useful binder must not only grip its target but also avoid everything else in a cell, and published papers often acknowledge that proteome-wide specificity testing was not performed. [16]

Reported numbers deserve care. Many headline hit rates come from the developers of the tool being measured, on targets of their choosing, with no replication. Two 2025 papers in Science found, from different directions, that AI-redesigned proteins frequently match a predicted structure while losing the function that mattered. [36][37]

Finally, structure is not biology. A design that folds correctly can still be immunogenic, unstable in serum, hard to manufacture, or simply ineffective in an animal, and none of those properties is well predicted by the models now in use.

Biosecurity and dual use

Design tools that can produce a new binder on request can also, in principle, be pointed at harmful targets, and the field has treated this as its own problem rather than waiting for regulators.

On March 8, 2024, researchers published a set of voluntary commitments under the banner Responsible AI x Biodesign, growing out of a summit held at the Institute for Protein Design in October 2023. Signatories commit as individuals, not institutions, to ten points: pursuing benefit-oriented research, supporting outbreak preparedness, buying synthetic DNA only from providers that screen orders to industry standards, helping improve that screening, continuously evaluating design software for risk, improving risk-evaluation methods, holding regular secure review meetings, reporting concerning practices, communicating both benefits and risks, and revising and spreading the principles. The statement had more than 90 signatories at launch and 192 as of October 2026. [38]

The sharpest test of those commitments came from the screening side. In a paper published in Science on October 2, 2025, Bruce Wittmann, Eric Horvitz and colleagues, working at Microsoft Research with the International Gene Synthesis Consortium and others, reported that the software used to screen commercial DNA synthesis orders reliably flagged wild-type sequences of concern but did not reliably flag variants redesigned with generative protein tools. The work was entirely computational: no proteins were made. Before publication the authors disclosed the finding to the United States government and to the developers of the screening software, patches were developed and deployed, and detection improved substantially, though the authors and outside commentators noted that gaps remain. The team described the problem as a biosecurity zero-day, borrowing the software security term. [36] A companion Science paper published on June 20, 2025, led by researchers at the National Institute of Standards and Technology using harmless proxy proteins, found the other half of the picture: current tools can generate variants that match a target's predicted structure but do not reliably preserve its function, so rewriting a sequence while keeping activity remains difficult and resource-intensive. [37]

Policy has moved less smoothly. Executive Order 14110, signed in October 2023, directed federal funders to require that federally funded researchers buy synthetic nucleic acids only from providers following a screening framework, and the Office of Science and Technology Policy released that framework on April 29, 2024. [39] The order was revoked in January 2025, and Executive Order 14292, "Improving the Safety and Security of Biological Research," signed on May 5, 2025, paused the 2024 framework and gave OSTP 90 days to revise or replace it, along with a 180-day deadline for a strategy to extend verifiable screening beyond federally funded work. [40] As of mid-2026 no replacement framework had been published; a June 24, 2026 letter from Senator Jon Ossoff to OSTP noted that more than 400 days after the order it was unclear whether a revised framework existed. [41]

References

The Royal Swedish Academy of Sciences. (2024, October 9). Press release: The Nobel Prize in Chemistry 2024. https://www.nobelprize.org/prizes/chemistry/2024/press-release/
Abbass, J., & Nebel, J.-C. (2020). Rosetta and the Journey to Predict Proteins' Structures, 20 Years on. Current Bioinformatics, 15(6), 611-628. https://doi.org/10.2174/1574893615999200504103643
Rosetta@home. Baker Laboratory, University of Washington. https://boinc.bakerlab.org/
University of Washington News. Video gamers design brand new proteins. https://newsroom.uw.edu/news/video-gamers-design-brand-new-proteins
Kuhlman, B., Dantas, G., Ireton, G. C., Varani, G., Stoddard, B. L., & Baker, D. (2003). Design of a novel globular protein fold with atomic-level accuracy. Science, 302(5649), 1364-1368. https://doi.org/10.1126/science.1089427
Rothlisberger, D., Khersonsky, O., Wollacott, A. M., et al. (2008). Kemp elimination catalysts by computational enzyme design. Nature, 453, 190-195. https://doi.org/10.1038/nature06879
Silva, D.-A., Yu, S., Ulge, U. Y., et al. (2019). De novo design of potent and selective mimics of IL-2 and IL-15. Nature, 565, 186-191. https://doi.org/10.1038/s41586-018-0830-7
Institute for Protein Design. COVID-19 vaccine with IPD nanoparticles wins full approval abroad. University of Washington. https://www.ipd.uw.edu/?p=6776
UW Medicine Newsroom. UW Medicine COVID-19 vaccine wins South Korea approval. https://newsroom.uw.edu/news-releases/uw-medicine-covid-19-vaccine-wins-south-korea-approval
Dauparas, J., Anishchenko, I., Bennett, N., et al. (2022). Robust deep learning-based protein sequence design using ProteinMPNN. Science, 378(6615), 49-56. https://doi.org/10.1126/science.add2187
Bennett, N. R., Coventry, B., Goreshnik, I., et al. (2023). Improving de novo protein binder design with deep learning. Nature Communications, 14, 2625. https://doi.org/10.1038/s41467-023-38328-5
Jumper, J., Evans, R., Pritzel, A., et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596, 583-589. https://doi.org/10.1038/s41586-021-03819-2
Baek, M., DiMaio, F., Anishchenko, I., et al. (2021). Accurate prediction of protein structures and interactions using a three-track neural network. Science, 373(6557), 871-876. https://doi.org/10.1126/science.abj8754
Watson, J. L., Juergens, D., Bennett, N. R., et al. (2023). De novo design of protein structure and function with RFdiffusion. Nature, 620(7976), 1089-1100. https://doi.org/10.1038/s41586-023-06415-8
Ingraham, J. B., Baranov, M., Costello, Z., et al. (2023). Illuminating protein space with a programmable generative model. Nature, 623(7989), 1070-1078. https://doi.org/10.1038/s41586-023-06728-8
Zambaldi, V., La, D., Chu, A. E., et al. (2024). De novo design of high-affinity protein binders with AlphaProteo. arXiv preprint arXiv:2409.08022. https://arxiv.org/abs/2409.08022
EvolutionaryScale. (2024, June 25). ESM3: Simulating 500 million years of evolution with a language model. https://www.evolutionaryscale.ai/blog/esm3-release
Hayes, T., Rao, R., Akin, H., et al. (2025). Simulating 500 million years of evolution with a language model. Science, 387(6736), 850-858. https://doi.org/10.1126/science.ads0018
Lauko, A., Pellock, S. J., Sumida, K. H., et al. (2025). Computational design of serine hydrolases. Science, 388, eadu2454. https://doi.org/10.1126/science.adu2454
Institute for Protein Design. (2025, December). RFdiffusion2: efficient enzymes from prompts. University of Washington. https://www.ipd.uw.edu/2025/12/rfdiffusion2-efficient-enzymes-from-prompts
Institute for Protein Design. (2025, December 4). RFdiffusion3 now available. University of Washington. https://www.ipd.uw.edu/2025/12/rfdiffusion3-now-available
MIT CSAIL. (2025, June). MIT releases breakthrough protein-binding affinity model, expanding role of AI in drug discovery. https://www.csail.mit.edu/news/mit-releases-breakthrough-protein-binding-affinity-model-expanding-role-ai-drug-discovery
Abramson, J., Adler, J., Dunger, J., et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 630(8016), 493-500. https://doi.org/10.1038/s41586-024-07487-w
Pacesa, M., Nickel, L., Schellhaas, C., et al. (2025). One-shot design of functional protein binders with BindCraft. Nature, 646, 483-492. https://doi.org/10.1038/s41586-025-09429-6
Chai Discovery. (2025, June 30). Introducing Chai-2: zero-shot antibody design in a 24-well plate. https://www.chaidiscovery.com/news/introducing-chai-2
Generate Biomedicines. Pipeline and corporate updates, including the GB-0895 Phase 3 SOLAIRIA program. https://generatebiomedicines.com/?p=468
Xaira Therapeutics. (2024, April 23). Xaira Therapeutics launches to deliver transformative medicines by advancing and harnessing AI for drug discovery and development. Business Wire. https://www.businesswire.com/news/home/20240423707240/en/Xaira-Therapeutics-Launches-to-Deliver-Transformative-Medicines-by-Advancing-and-Harnessing-AI-for-Drug-Discovery-and-Development
TechCrunch. (2024, June 25). EvolutionaryScale, backed by Amazon and Nvidia, raises $142M for protein-generating AI. https://techcrunch.com/2024/06/25/evolutionaryscale-backed-by-amazon-and-nvidia-raises-142m-for-protein-generating-ai/
Gunderson Dettmer. (2025, November). EvolutionaryScale team joins Biohub in strategic transaction with Chan Zuckerberg Initiative. https://www.gunder.com/en/news-insights/client-news/gunderson-client-evolutionaryscale-team-joins-biohub-in-strategic-transaction-with-chan-zuckerberg-initiative
Isomorphic Labs. (2025, March 31). Isomorphic Labs announces $600m external investment round. https://www.isomorphiclabs.com/articles/isomorphic-labs-announces-600m-external-investment-round
BioSpace. (2022, November 14). Neoleukin Therapeutics announces third quarter 2022 financial results and corporate update. https://www.biospace.com/neoleukin-therapeutics-announces-third-quarter-2022-financial-results-and-corporate-update
Absci. (2025, November 13). Absci reports business updates and third quarter 2025 financial results. https://investors.absci.com/news-releases/news-release-details/absci-reports-business-updates-and-third-quarter-2025-financial/
Investing.com. (2026, January). Google-backed AI drug discovery startup Isomorphic Labs delays clinical trial timeline. https://www.investing.com/news/stock-market-news/googlebacked-ai-drug-discovery-startup-isomorphic-labs-delays-clinical-trial-timeline-4456134
Protein Structure Prediction Center. CASP16 (2024). https://predictioncenter.org/casp16/
ProtDBench. (2026). Benchmark of generative protein design methods, International Conference on Machine Learning. arXiv preprint arXiv:2605.04118. https://arxiv.org/abs/2605.04118
Wittmann, B. J., Alexanian, T., Bartling, C., et al. (2025). Strengthening nucleic acid biosecurity screening against generative protein design tools. Science, 390(6768), 82-87. https://doi.org/10.1126/science.adu8578
Ikonomova, S., Wittmann, B., Piorino, F., et al. (2025, June 20). Experimental evaluation of AI-driven protein design risks using safe biological proxies. Science. https://www.nist.gov/publications/experimental-evaluation-ai-driven-protein-design-risks-using-safe-biological-proxies
Responsible AI x Biodesign. (2024, March 8). Community values, guiding principles, and commitments for the responsible development of AI for protein design. https://responsiblebiodesign.ai/
Office of Science and Technology Policy. (2024, April 29). Framework for Nucleic Acid Synthesis Screening. https://www.aspr.gov/S3/Pages/OSTP-Framework-for-Nucleic-Acid-Synthesis-Screening.aspx
Executive Order 14292. (2025, May 5). Improving the Safety and Security of Biological Research. https://www.govinfo.gov/content/pkg/DCPD-202500561/html/DCPD-202500561.htm
Office of Senator Jon Ossoff. (2026, June 24). Letter to the Office of Science and Technology Policy on nucleic acid synthesis screening. https://www.ossoff.senate.gov/wp-content/uploads/2026/07/26.06.24_OSTP-AI-Letter.pdf

Editing: Protein design