Future Horizons

The Synthetic Biology Convergence: How AI Is Becoming the Master Architect of Life

AlphaGenome reads a million base pairs at a time, Evo 2 writes genomes, RFdiffusion invents enzymes evolution never tried. What AI can now design in biology, and what still has to survive a wet lab.

Gabriele Masetti · 2026-08-21

A discipline that used to read is learning to write

For most of its history, structural biology was a discipline of interpretation. Scientists crystallized proteins, fired X-rays through them, and spent months or years inferring three-dimensional shapes from diffraction patterns. Understanding how a single protein folded could be the substance of an entire career. That changed with the arrival of deep-learning structure predictors, and it changed again, more quietly but more consequentially, when the same underlying techniques were turned around and pointed at generation rather than prediction.

The distinction matters because it marks the difference between AI as a research instrument and AI as something closer to a co-author of biological form. Reading a genome or a fold is an act of interpretation; writing a new protein, enzyme, or genome is an act of authorship. That shift from reading life to writing it is the actual substance of what gets loosely called the "synthetic biology convergence," and it is now advanced enough that a Nobel Prize has already been awarded for it.

From prediction to design: the Nobel Prize as inflection point

The 2024 Nobel Prize in Chemistry made the field's turning point official. The prize was split, with one half going to David Baker of the University of Washington's Institute for Protein Design "for computational protein design," and the other half jointly to Demis Hassabis and John Jumper of Google DeepMind "for protein structure prediction." That the committee drew this exact distinction is instructive: it recognized prediction and design as two related but separate scientific achievements, occurring in close succession and building on shared computational lineage.

The prediction side of that lineage is well known. AlphaFold2 demonstrated that a deep neural network could predict a protein's three-dimensional structure from its amino acid sequence with an accuracy rivaling experimental methods, collapsing a problem that had taken structural biologists decades to chip away at. DeepMind and its sister company Isomorphic Labs followed in May 2024 with AlphaFold3, published in Nature, which extended the same diffusion-based architecture beyond single proteins to model complexes involving DNA, RNA, small-molecule ligands, ions, and modified residues.

According to DeepMind's own published results, AlphaFold3 achieved roughly a 50 percent improvement in accuracy for protein-molecule interaction predictions over prior methods, and in some categories it doubled accuracy outright. Isomorphic Labs has since built on that foundation with what it calls a Drug Design Engine, using the model in partnership with pharmaceutical companies to pursue previously intractable drug targets. Its pipeline remains pre-clinical in public: as of September 2026 no trial registered on ClinicalTrials.gov names the company as sponsor.

Baker's contribution runs in the opposite direction. His lab's tool RFdiffusion, built on an earlier structure-prediction system called RoseTTAFold, does not predict how an existing sequence folds; it generates entirely new protein backbones that have never existed in nature, using a denoising diffusion process that starts from noise and is gradually refined toward a structure with a desired shape or function.

In effect, if AlphaFold answers "what shape is this sequence," RFdiffusion answers "what sequence would produce this shape," inverting the problem. The Institute for Protein Design has used the approach and related methods to generate novel proteins with applications spanning drug candidates, vaccines, nanomaterials, and biosensors that have no precedent in biological evolution.

The generative turn made concrete

The clearest illustration of what generative protein design can do predates the Nobel by a year but anticipates exactly the trajectory the prize recognized. In 2023, postdoctoral researchers Andy Hsien-Wei Yeh and Christoffer Norn in Baker's lab published a Nature paper describing the de novo design of luciferases, the light-emitting enzymes familiar from fireflies, using a deep-learning strategy they called "family-wide hallucination."

The resulting enzymes were small, at roughly 13.9 kilodaltons, remarkably thermostable above 95 degrees Celsius, and catalytically comparable to natural luciferases, despite having no evolutionary history at all. It is one thing to predict how a natural protein folds; it is another to generate, entirely computationally, a functional enzyme with properties superior to its natural counterparts in some respects.

That kind of workflow has continued to mature. A 2024 pipeline called GRACE combined RFdiffusion for backbone generation, the sequence-design tool ProteinMPNN, and molecular dynamics simulation for screening, and used it to design functional carbonic anhydrase enzymes, identifying working candidates out of a pool of roughly ten thousand generated sequences.

Separately, a team of former Meta AI researchers who left to found the company EvolutionaryScale released ESM3, a protein language model trained on the sequence, structure, and function of more than 2.7 billion proteins, which its creators describe as the first generative model to reason jointly over all three properties at once.

EvolutionaryScale demonstrated ESM3 by generating a novel fluorescent protein whose sequence was only 58 percent similar to its nearest known natural relative, describing the result as equivalent to roughly 500 million years of simulated evolution compressed into a single generative step. The company raised a $142 million seed round from investors including Lux Capital, Amazon Web Services, Nat Friedman, Daniel Gross, and Nvidia's venture arm.

Genomic foundation models: designing at the scale of whole genomes

Protein design operates on molecules. A newer and arguably more consequential wave of models operates on genomes, and two of them define the current state of the field.

Google DeepMind's AlphaGenome takes the reading half. Announced on 25 June 2025 and published in Nature on 28 January 2026, it ingests up to a million base pairs of DNA at once and predicts thousands of molecular properties from that sequence alone: where genes start and stop in a given cell type, how RNA is spliced, how much of it is made, which stretches of DNA sit open, where proteins bind. DeepMind reports it outperformed external models on 22 of 24 sequence-prediction evaluations and matched or beat the best of them on 24 of 26 variant-effect evaluations.

The machinery exists for the large majority of the genome that codes for no protein. A variant in a regulatory region changes no amino acid, so the protein-centred scoring tools geneticists lean on have nothing to say about it; AlphaGenome scores it by predicting what the surrounding sequence does. A model that generates nothing belongs in an essay about writing for a plain reason — nobody writes regulatory DNA they cannot read.

Evo 2 takes the writing half. In February 2025 the nonprofit Arc Institute released it as a preprint, with the peer-reviewed paper following in Nature on 4 March 2026. It was trained, according to Arc, on more than 9.3 trillion nucleotides drawn from the genomes of more than 128,000 species spanning all domains of life, using a novel architecture called StripedHyena 2 and roughly 40 billion parameters, built on Nvidia's DGX Cloud infrastructure.

Where a protein language model reasons over folded chains of amino acids, Evo 2 reasons over raw DNA sequence, and it can process sequence windows up to a million nucleotides long, long enough to capture relationships between genomic elements separated by large distances, something previous genomic models could not do.

The capabilities Arc reports are consequential in both directions. Evo 2 can be used analytically, to identify disease-causing mutations in human genes by scoring how disruptive a given variant is likely to be. But it is also, per its own description, a generative model: capable of designing new genetic sequences as long as the genomes of simple bacteria. Its predecessor, the original Evo model, worked only with prokaryotic genomes; Evo 2 was extended to train on eukaryotic sequences as well, including human DNA, specifically to generalize across the tree of life rather than a narrow slice of it.

Read together, AlphaFold3, RFdiffusion, ESM3, AlphaGenome, and Evo 2 describe a stack rather than a set of unrelated projects: a structure-prediction layer that tells you what a molecule will look like, a protein-design layer that generates molecules to specification, a genomic layer that reasons over and generates entire sequences of DNA, and, increasingly, drug-design layers built on top of all of it. That stack is what justifies calling the moment a convergence. Structural biology, protein engineering, and genomics are no longer three separate disciplines borrowing occasional computational tools from one another; they are converging on a shared generative modeling substrate.

The design-build-test loop gets compressed

What changes when this stack matures is not simply that biologists can do old things faster. It is that the design-build-test cycle that has always constrained biological engineering — propose a hypothesis, synthesize it, test it in the lab, iterate — gets compressed at its most expensive stage. Historically, generating a candidate protein or genetic circuit worth testing required either painstaking rational design grounded in mechanistic understanding, or brute-force screening of large random or natural libraries.

Generative models offer a third path: computationally propose large numbers of plausible, purpose-built candidates before any wet-lab work begins, then synthesize and test only the most promising few. The GRACE enzyme pipeline's ability to narrow ten thousand computational candidates down to working carbonic anhydrase variants is a small-scale demonstration of exactly that compression. Whether it holds at industrial scale, across harder design targets than fluorescent proteins and enzymes with well-characterized active sites, remains an open empirical question, but the direction of travel is unambiguous.

The shadow side: biosecurity catches up to capability

The same properties that make generative biology powerful for beneficial research — the ability to propose novel biological sequences without needing to understand every mechanistic detail in advance — are dual-use by construction, and the field's own institutions have started to say so plainly rather than treat it as a hypothetical. RAND has published some of the most concrete assessment work in this area.

A RAND Perspective examined whether general-purpose foundation models could meaningfully lower the barrier to biological weapons development, and reported that the frontier systems it tested, specifically Llama 3.1 405B, GPT-4o, and Claude 3.5 Sonnet, were able to produce accurate instructions for recovering live poliovirus from synthetic DNA when prompted, a test case RAND treats as a proxy for the broader risk of reconstructing dangerous pathogens from published genetic sequence data.

Separately, RAND has built a Global Risk Index for AI-enabled Biological Tools, a structured framework assessing tools by capability, potential for misuse, accessibility, and technical maturity; in a 2025 assessment of 57 state-of-the-art tools, the index flagged 13 as requiring immediate mitigating action and another 15 as warranting case-by-case follow-up review.

Policymakers have moved, albeit unevenly. On May 5, 2025, the U.S. executive branch signed an Executive Order on "Improving the Safety and Security of Biological Research," which requires a verifiable framework for screening nucleic acid synthesis orders and ties compliance to eligibility for federal life-science research funding, with violations risking both immediate funding revocation and up to five years of ineligibility.

That order builds on a screening gap biosecurity researchers had flagged for years: most, though not all, commercial DNA synthesis providers already screen orders voluntarily through membership in the industry-led International Gene Synthesis Consortium, but participation has never been mandatory, and the same synthesis capability that lets a Baker-lab postdoc order a novel luciferase gene also lets a bad actor order a sequence pulled from a published pathogen genome.

In January 2026, the U.S. Senate introduced the Biosecurity Modernization and Innovation Act, which would direct the Commerce Department to formalize binding nucleic acid synthesis security regulations rather than rely on industry self-policing. It has not been enacted; the binding requirement still runs through the executive order's funding condition rather than through statute.

The honest way to characterize this landscape is not that AI-designed biology is on the verge of producing novel bioweapons from scratch — the documented, verified risk concentrates on AI models lowering the knowledge barrier to reconstructing or handling known dangerous pathogens, which is a narrower and more tractable problem than de novo weapon design. But it is also not a manufactured panic; RAND's own researchers, working through structured evaluation rather than speculation, concluded that today's frontier models measurably assist with specific, dangerous, previously hard-to-access technical steps.

Screening at the DNA synthesis chokepoint, precisely because it sits downstream of both natural and AI-generated sequences alike, is the control policymakers have converged on as the most tractable point of intervention, since it does not require distinguishing which sequences originated from a generative model and which came from a human researcher's memory of the literature.

What "master architect" should and shouldn't mean

It is tempting, and increasingly common in coverage of this space, to describe AI as though it has become an autonomous designer of life, quietly running experiments no human oversees. That framing overstates the current reality. Every system discussed here, from RFdiffusion to Evo 2 to ESM3, produces candidates that still require physical synthesis and empirical validation in a wet lab; a generative model proposing a protein sequence is a hypothesis generator, not an organism-building assembly line, and the gap between a computationally plausible design and a functioning biological system that survives contact with an actual cell or organism remains large and, in most cases, still measured in months of lab work rather than the seconds it takes the model to generate the candidate.

The gauntlet is survivable, and the clearest evidence is a drug. Rentosertib, a TNIK inhibitor that Insilico Medicine's generative models proposed for idiopathic pulmonary fibrosis, published Phase IIa results in Nature Medicine on 3 June 2025 and opened a 320-patient Phase III trial on 7 July 2026. Generation was the fast part; everything after it ran at the speed of cell cultures, ethics committees and regulators.

What has genuinely changed, and what earns the "master architect" framing, is where the intellectual center of gravity in biological design now sits. A decade ago, the rate-limiting step in engineering a new protein or genetic sequence was a scientist's mechanistic intuition about how amino acids or nucleotides would behave in combination. Today, for functions that have well-characterized structural targets, the rate-limiting step is increasingly the wet-lab cycle of synthesizing and testing what a generative model has already proposed.

That inversion is exactly why a Nobel Prize went to computational design in the same year the same committee honored structure prediction, and exactly why biosecurity researchers, rather than dismissing the technology as hype, have spent 2025 building risk indices and pushing binding synthesis-screening law.

The architecture of life is not yet fully in AI's hands, but the blueprint stage increasingly is, and the institutions built to catch dangerous blueprints before they reach a lab bench are still being assembled in real time, running roughly a year or two behind the capability they are meant to govern.