AI-Designed DNA: How Neural Networks Build New Proteins

In January 2026, a biotech company called Generate:Biomedicines dosed the first Phase 3 patient with an antibody drug, GB-0895, that a generative AI model designed from scratch to treat severe asthma. No human chemist drew that molecule. A neural network trained on protein sequences proposed it, a lab built it, and it is now working its way through the same regulatory gauntlet as any other drug. This is not a hypothetical. AI-designed molecules are already inside human bodies. This piece explains the mechanics behind that shift: how models like AlphaFold, RFdiffusion, ESM3, and Evo turned DNA and protein design from a slow trial-and-error craft into something closer to computer-aided engineering, and where the approach still runs into hard biological limits.

Educational content, not medical advice. This article explains research and technology trends. It is not a recommendation for any treatment, product, or clinical decision. Talk to a qualified physician about any medical question.

Why DNA design used to be a brute-force problem

Designing a new protein or genetic circuit has traditionally worked the way locksmithing worked before computer-aided design: cut a key, test it in the lock, file it down, test again. Every round costs time and money, and the number of possible designs is enormous. A modest protein of 300 amino acids has roughly 20³⁰⁰ possible sequences, a number that dwarfs the count of atoms in the observable universe. Nature only ever explored a sliver of that space through billions of years of evolution.

Before machine learning entered the picture, protein engineers relied on directed evolution: mutate a gene randomly, screen thousands of variants for the property you want (say, heat tolerance or higher enzyme activity), keep the best ones, and repeat. Frances Arnold won a share of the 2018 Nobel Prize in Chemistry for pioneering this method, and it remains one of the most reliable tools in the field (Nobel Prize, 2018). But directed evolution is slow. Each round of mutation and screening in a lab can take weeks, and researchers are still searching blindly around whatever starting point they picked.

The other bottleneck was structure. To design a protein rationally, rather than by trial and error, you need to know how its amino acid chain folds into a three-dimensional shape, because shape determines function. Predicting that fold from sequence alone was considered one of biology's hardest open problems for over 50 years, formalized as the "protein folding problem" and turned into a standing competition (CASP, the Critical Assessment of Structure Prediction) in 1994. For decades, progress was incremental. That changed in 2020.

A modest 300-amino-acid protein has about 20³⁰⁰ possible sequences — a number that dwarfs the estimated 10⁸² atoms in the observable universe. AI models don't search this space blindly; they've learned the statistical patterns evolution already found.

How AlphaFold cracked structure prediction, and what came after

DeepMind's AlphaFold2, released in 2020 and detailed formally in 2021, solved protein structure prediction with startling accuracy, hitting roughly 90% accuracy within about 1 angstrom (0.1 nanometer) of experimentally measured structures at the 2020 CASP14 competition (Jumper et al., 2021, Nature). By 2022, DeepMind had released predicted structures for over 200 million proteins, covering nearly every protein sequence known to science (EMBL-EBI AlphaFold Protein Structure Database). In October 2024, the Royal Swedish Academy of Sciences awarded the Nobel Prize in Chemistry jointly to Demis Hassabis and John Jumper of DeepMind for AlphaFold, and to David Baker of the University of Washington for computational protein design (Nobel Prize press release, 2024).

AlphaFold answers one direction of the problem: given a sequence, what shape does it fold into? Baker's lab built the tools that answer the opposite question, the one that actually matters for synthetic biology: given a shape or function you want, what sequence produces it? That's called protein design, and it's the more useful skill for building new medicines, enzymes, and materials.

Baker's lab released RFdiffusion in 2023, a generative diffusion model, the same general technique behind image generators like Midjourney, adapted to build protein backbones. Diffusion models work by starting with noise and gradually removing it in a way that's guided toward a target, whether that target is a coherent image or a stable, functional protein shape. RFdiffusion generates protein backbones, including shapes never observed in any natural organism, and can then have a sequence design step fill in which amino acids should sit at each position (Watson et al., 2023, Nature).

The efficiency gain is the headline result. With earlier design methods, researchers sometimes had to synthesize and test tens of thousands of candidate proteins to find even one that worked as intended. With RFdiffusion, the Baker lab's team reported success with as few as one candidate tested per design challenge in several cases (Baker Lab, 2023). That's not a modest improvement; it changes what's economically feasible to attempt.

How does AI actually generate a new protein sequence?

Two complementary approaches dominate. Structure-first tools like RFdiffusion sculpt a 3D backbone shape, then a separate network (such as ProteinMPNN) fills in which amino acids stabilize that exact shape. Sequence-first tools like ESM3 treat protein sequences the way a large language model treats sentences, predicting which amino acid comes next based on patterns learned from hundreds of millions of natural proteins, then extending that skill to generate entirely new ones.

ESM3, Evo, and the rise of biological language models

If AlphaFold and RFdiffusion are structure-focused, a parallel line of research treats biological sequences the way ChatGPT treats English text: as a language to be modeled statistically. The idea is that evolution has already run an enormous, multi-billion-year experiment, and the patterns left behind in DNA and protein sequences encode deep information about what works.

EvolutionaryScale, a company founded by former Meta AI researchers, released ESM3 in mid-2024, a multimodal model trained on 2.78 billion protein sequences and 236 million known structures, scaling up to 98 billion parameters in its largest version (EvolutionaryScale, 2024). ESM3 doesn't just predict structure from sequence; it can generate new sequences given a target structure, function, or combination of both. In a widely reported demonstration, the researchers prompted ESM3 to generate a new fluorescent protein and produced one dubbed esmGFP, whose sequence differs from any known natural fluorescent protein by about 58%, a degree of divergence roughly comparable to what would be expected from 500 million years of natural evolution along a single lineage (Hayes et al., 2025, Science).

A separate effort out of the Arc Institute, Stanford, and UC Berkeley pushed the language-model approach down to raw DNA rather than protein sequence. Evo, released in November 2024, is a 7-billion-parameter model trained on prokaryotic and phage genomes at single-nucleotide resolution across a context window of 131,000 tokens, letting it reason across sequence stretches that span whole genes and their regulatory regions at once (Nguyen et al., 2024, Arc Institute). Its 2025 successor, Evo 2, scaled to 40 billion parameters trained on more than 9.3 trillion nucleotides drawn from over 128,000 whole genomes spanning bacteria, archaea, and eukaryotes including humans and plants (Arc Institute, 2025). Evo 2 can both predict the likely effect of a genetic mutation and generate new DNA sequences, including, in early demonstrations, sequences resembling functional genome fragments.

What's the difference between AlphaFold and RFdiffusion?

AlphaFold predicts the 3D structure of a protein when you already know its amino acid sequence. RFdiffusion works in the opposite direction: it generates new protein structures, including ones with no natural counterpart, that you specify by function or shape, and then a companion model fills in a sequence that would fold into that structure.

From lab bench to genetic circuit: designing more than single proteins

Individual proteins are the simplest unit synthetic biologists design, but the field's real ambition is bigger: engineering whole genetic circuits, systems of genes that switch each other on and off like logic gates, and full metabolic pathways that turn cheap feedstock into a valuable product.

A well-cited example predates the newest generative models but illustrates the target economics AI is now accelerating: in 2013, Jay Keasling's group at UC Berkeley engineered yeast to produce artemisinic acid, the precursor to the antimalarial drug artemisinin, via a bioengineered metabolic pathway, work that was later licensed for industrial-scale semi-synthetic artemisinin production, reducing dependence on Artemisia annua plant extraction, which is slow and subject to volatile harvests (Paddon et al., 2013, Nature). That pathway took years of manual metabolic engineering to optimize. Machine learning models trained on sequence-to-expression data now aim to compress that iteration cycle by predicting, before anything is synthesized, which promoter and ribosome-binding-site combinations will hit a target expression level.

Genetic circuits work similarly to digital logic. A circuit might be designed so that Gene C only switches on when both Gene A and Gene B are active, an AND gate built out of biology instead of silicon. The foundational framework for this kind of circuit design was laid out by synthetic biology researchers Ahmad Khalil and James Collins, whose review of synthetic gene circuit design remains a standard reference in the field (Khalil & Collins, 2010, Nature Reviews Genetics). What's changed since is that instead of hand-tuning each part through repeated lab cycles, researchers increasingly train models on large datasets of sequence-to-behavior measurements and use those models to rank candidate designs before touching a pipette.

The pattern across every one of these tools is the same: AI doesn't replace the wet lab. It shrinks the number of expensive physical experiments needed to find something that works, by ranking candidates computationally first.

AI-designed antibodies are already in human trials

The clearest real-world proof that this technology works is happening in drug development, where the stakes and the scrutiny are highest. Generate:Biomedicines designs therapeutic proteins, including antibodies, using generative models trained on large protein sequence datasets rather than starting from a natural antibody and modifying it. Its lead program, GB-0895, an anti-TSLP antibody engineered for severe asthma and other respiratory disease, began dosing its first Phase 3 patient in January 2026, reportedly the first AI-designed antibody to reach that stage of human testing (GenEngNews, 2026). The company's separate program, GB-0669, an AI-generated antibody targeting a viral target, has reported Phase 1 safety and pharmacokinetic data in healthy volunteers.

Xaira Therapeutics launched in April 2024 with more than $1 billion in committed funding, one of the largest initial biotech financings on record, built explicitly around applying diffusion models (the same generative technique behind RFdiffusion and image generators) to biologic drug design, including antibodies (Nature, 2024). Absci, another AI-drug-design company, has moved its TL1A antibody candidate ABS-101 and its androgenetic alopecia candidate ABS-201 into Phase 1/2a trials.

None of this means AI-designed drugs are guaranteed to succeed. Reaching Phase 3 means a drug has cleared safety and preliminary efficacy hurdles, not that it will ultimately win regulatory approval; most drug candidates that reach Phase 3 still fail there or in the approval process. What's notable is narrower and more concrete: AI-generated molecular designs are clearing the same early bars that traditionally-designed drugs clear, at a pace and hit-rate that has attracted serious institutional investment.

Where AI-driven synthetic biology still breaks down

It's worth being direct about the limits, because the marketing language around this field often outruns what the models can reliably do.

Structure prediction confidence varies enormously by protein type. AlphaFold reports a per-residue confidence score (pLDDT), and it is far more reliable for stable, well-folded domains than for intrinsically disordered regions of proteins, which make up a substantial share of the human proteome and don't have a single fixed shape. Predicting how a protein behaves inside a living cell, surrounded by thousands of other interacting molecules, is a much harder problem than predicting its shape in isolation.

Generated sequences are hypotheses, not guarantees. A model like ESM3 or RFdiffusion proposes candidate designs it calculates are plausible; it does not certify that a candidate will actually express, fold correctly, and function once synthesized in a real cell. Every credible synthetic biology program still validates computational designs experimentally before moving forward, exactly because model output and biological reality diverge often enough to matter.

Context effects don't transfer well between organisms. A genetic promoter that reliably drives strong expression in E. coli can behave completely differently in yeast or mammalian cells, because the transcription and translation machinery differs. Models trained primarily on bacterial data need retraining, or at least careful validation, before their predictions can be trusted in a different host organism.

Engineered organisms can lose their engineered traits over time. If a gene inserted into yeast carries any metabolic cost, cells that lose or silence that gene through spontaneous mutation often grow slightly faster than engineered cells that keep it. Over many generations, that selection pressure can erode the population's engineered function, which is why synthetic biologists build in genetic stabilizers or accept that production strains need periodic renewal.

Regulatory approval is slow regardless of how the molecule was designed. An AI-designed antibody still has to clear the same FDA (or EMA, or other regulator) pathway as any other biologic: preclinical toxicology, then Phase 1, 2, and 3 trials, typically spanning many years total. AI can shrink the design phase from months to weeks; it does not shrink the clinical testing timeline, which exists to protect patient safety and is not primarily a discovery bottleneck.

How AI-driven design compares to older synthetic biology methods

Approach	Core method	Typical iteration cost	Best suited for
Directed evolution	Random mutation + screening of thousands of variants	Weeks per round, many rounds needed	Improving an existing protein's stability or activity
Rational design (pre-AI)	Manual sequence/structure reasoning by expert biochemists	Months, heavy expert labor	Small, well-understood modifications
AlphaFold-class structure prediction	Deep learning predicts 3D shape from sequence	Minutes to hours, computational only	Understanding an existing or proposed protein's likely shape
RFdiffusion-class generative design	Diffusion model generates novel backbone + sequence	Days (design) plus lab validation	Creating proteins with no natural equivalent, from scratch
ESM3/Evo-class language models	Transformer trained on sequence data predicts/generates sequences	Minutes to hours, computational only	Generating sequence variants, predicting mutation effects

Table: How AI-driven protein and DNA design methods compare to traditional synthetic biology approaches. Iteration costs are approximate and vary by project complexity.

AI and CRISPR are solving different problems

It's easy to conflate AI-driven protein design with CRISPR gene editing because both show up in headlines about "engineering biology," but they answer different questions. CRISPR is a tool for editing DNA that already exists inside a living cell or organism, cutting out a faulty sequence and replacing it with a corrected one. It's the technology behind Casgevy, the first CRISPR-based therapy approved by the FDA in December 2023 for sickle cell disease. For the mechanics of how CRISPR itself works and its clinical track record, see The CRISPR Gene Editing Breakthrough.

AI-driven synthetic biology, by contrast, is mostly about designing entirely new proteins or genetic circuits before any cell is involved, work that happens computationally first. The two fields increasingly overlap: AI models are now used to help predict which CRISPR guide RNA sequences will cut precisely at the intended DNA location and avoid off-target edits elsewhere in the genome, since off-target cutting is one of CRISPR's most persistent safety concerns. In that sense, AI is becoming a design layer that sits on top of CRISPR, not a replacement for it.

What this means

The practical shift is this: designing a new protein or genetic circuit used to mean picking a starting point from nature and nudging it, because searching the full space of possible sequences by hand was hopeless. AI models trained on the outcomes of billions of years of evolution now let researchers search that space computationally first, testing ideas in software before committing to the weeks and dollars a physical lab experiment costs. That doesn't eliminate the wet lab, the regulatory process, or the years of clinical testing a new drug still requires. What it changes is the funnel feeding into all of that, and the fact that a fully AI-designed antibody is now in Phase 3 human trials is the clearest sign yet that the funnel is producing real, testable biology, not just impressive demos.

Frequently asked questions

Q: Can AI actually design a working protein, or does it just predict structures?

A: Both, depending on the tool. AlphaFold predicts structure from a known sequence. Generative tools like RFdiffusion and ESM3 go further and create new sequences, including ones for proteins that don't exist in nature, that researchers then synthesize and test in a lab.

Q: Is any AI-designed drug actually approved for use yet?

A: Not as of mid-2026. Several AI-designed antibodies, including Generate:Biomedicines' GB-0895, have reached Phase 2 or Phase 3 human trials, but reaching full regulatory approval typically still takes additional years of clinical testing after that.

Q: How is Evo different from AlphaFold?

A: AlphaFold works on proteins and predicts 3D structure. Evo works directly on raw DNA sequence at single-nucleotide resolution and can both predict the effects of mutations and generate new DNA sequences, including across genome-scale stretches.

Q: Does AI-designed synthetic biology replace CRISPR gene editing?

A: No. CRISPR edits DNA that already exists inside a cell. AI protein and DNA design tools mostly create new sequences computationally before any cell is involved. Increasingly, AI models are also used to help choose safer, more precise CRISPR targets.

Q: Could an AI-designed organism be dangerous if it escaped a lab?

A: It's a recognized biosecurity concern, which is why engineered organisms intended for controlled environments are typically built with safeguards like metabolic dependencies on nutrients not found in nature, or genetic kill switches that trigger self-destruction outside defined lab conditions.

Q: How accurate is AlphaFold, really?

A: At the 2020 CASP14 competition, AlphaFold2 achieved a median accuracy within about 1 angstrom (0.1 nanometer) of experimentally determined structures for well-behaved, ordered protein regions, a level considered competitive with the accuracy of experimental methods themselves for many targets. Accuracy is lower for intrinsically disordered protein regions.

Q: Who won the Nobel Prize for this work?

A: The 2024 Nobel Prize in Chemistry went jointly to Demis Hassabis and John Jumper of Google DeepMind for AlphaFold, and to David Baker of the University of Washington for computational protein design, the work underlying tools like RFdiffusion.

Q: What is a protein language model?

A: It's a neural network, architecturally similar to large language models like GPT, trained not on English text but on millions of protein sequences, learning statistical patterns that let it predict, generate, or evaluate new sequences the same way a text model predicts the next word in a sentence.

Q: Why can't AI just design any protein a researcher wants?

A: Models are constrained by what they learned from training data drawn largely from existing biology, and by the fact that predicted designs are hypotheses that still need experimental validation. A generated sequence can look statistically plausible and still fail to fold, express, or function correctly once synthesized.

Sources

1. Nobel Prize. "The Nobel Prize in Chemistry 2018" (Frances Arnold, directed evolution). nobelprize.org
2. Jumper, J., et al. (2021). "Highly accurate protein structure prediction with AlphaFold." Nature. nature.com/articles/s41586-021-03819-2
3. EMBL-EBI. "AlphaFold Protein Structure Database." alphafold.ebi.ac.uk
4. Nobel Prize. "The Nobel Prize in Chemistry 2024" press release (Hassabis, Jumper, Baker). nobelprize.org
5. Watson, J. L., et al. (2023). "De novo design of protein structure and function with RFdiffusion." Nature. nature.com/articles/s41586-023-06415-8
6. Baker Lab. "RFdiffusion: A generative model for protein design." (2023). bakerlab.org
7. Hayes, T., et al. (2025). "Simulating 500 million years of evolution with a language model." Science. science.org/doi/10.1126/science.ads0018
8. EvolutionaryScale. "ESM3: Simulating 500 million years of evolution with a language model." (2024). evolutionaryscale.ai
9. Arc Institute. "Evo: DNA foundation modeling from molecular to genome scale." (2024). arcinstitute.org
10. Arc Institute. "Evo 2: AI model designs the genetic code across all domains of life." (2025). arcinstitute.org
11. Paddon, C. J., et al. (2013). "High-level semi-synthetic production of the potent antimalarial artemisinin." Nature. nature.com/articles/nature12051
12. Khalil, A. S., & Collins, J. J. (2010). "Synthetic biology: applications come of age." Nature Reviews Genetics. nature.com/articles/nrg2775
13. GenEngNews. "AI-Designed Antibodies Achieve Atomic Precision to Enhance Drug Discovery." (2026). genengnews.com
14. Nature. "Generative AI platforms drive drug discovery dealmaking." (2024). nature.com/articles/d43747-024-00084-w