In October 2024, the Royal Swedish Academy of Sciences announced that the Nobel Prize in Chemistry would be shared between three scientists: David Baker of the University of Washington, for his work in computational protein design, and Demis Hassabis and John Jumper of Google DeepMind, for developing AlphaFold, the AI system that solved one of biology’s most enduring challenges. The prize committee described their work as revealing “the secrets of proteins” — the molecular machines that govern virtually every biological process from metabolism to immunity to development.
The Nobel committee’s recognition marked something unusual: a prize for work that had happened primarily within the past five years, to researchers who were still actively extending their discoveries, using tools that were still improving. In recognition of its groundbreaking impact, David Baker received the 2024 Nobel Prize in Chemistry for his work in de novo protein design, highlighting its promise for transforming biologic development. The word “de novo” — from scratch, from nothing, without reference to any protein that has ever existed in nature — is the key. Baker’s lab had not just learned to predict the shapes of natural proteins. It had learned to design entirely new ones.
In 2025 and 2026, the consequences of that learning are beginning to reach patients.
What AlphaFold Actually Did
To understand why this matters, it helps to understand what protein structure prediction actually is and why it was so hard.
A protein is a chain of amino acids — typically between 100 and 1,000 of them, drawn from an alphabet of 20 — that folds into a precise three-dimensional shape determined by the physics of molecular interaction. The shape determines the function. A protein that folds into the wrong shape is either useless or dangerous — many diseases, from Alzheimer’s to Parkinson’s to prion disorders, involve proteins that fold incorrectly and aggregate in toxic ways.
For fifty years, determining the three-dimensional structure of a protein required labor-intensive experimental techniques — X-ray crystallography, cryo-electron microscopy, NMR spectroscopy — that could take months to years for a single protein. The Protein Data Bank, established in 1971, had accumulated approximately 200,000 structures by 2020 — representing the cumulative work of thousands of laboratories over five decades. Most of the roughly 200 million known proteins had no determined structure.
In 2020, AlphaFold 2 changed this in a single competitive demonstration. At the Critical Assessment of Protein Structure Prediction competition — a biennial challenge where research groups attempt to predict protein structures from sequence alone — AlphaFold 2 achieved accuracy comparable to experimental methods, on proteins it had never seen, in a matter of hours per structure. The result was described by the competition’s organizers as a solution to a fifty-year-old grand challenge.
The AlphaFold Protein Structure Database expanded to 214 million predicted structures, covering nearly all cataloged proteins known to science. In two years, the number of known protein structures increased by a factor of more than 1,000. The entire proteome of every organism that had ever been sequenced became structurally available to researchers.
AlphaFold 3: From Prediction to Drug Design
AlphaFold 3, released by Google DeepMind and Isomorphic Labs in May 2024, extended the system’s capabilities beyond single protein structures to the interactions between proteins, small molecules, nucleic acids, and other biological entities. Unlike its predecessors, AlphaFold 3 uses a novel diffusion-based architecture that allows it to model complex molecular systems beyond single proteins — accurately predicting binding sites and interaction energies, and enabling rational drug design and target identification.
This extension is what makes AlphaFold relevant to drug discovery rather than just biology. A drug works by binding to a protein target — typically fitting into a pocket on the protein’s surface and altering its function. Predicting how a candidate drug molecule will fit into its target, with sufficient accuracy to distinguish strong binders from weak ones, is one of the core computational problems in pharmaceutical development. Before AlphaFold 3, this required either expensive experimental measurement or computational methods that were unreliable enough to produce many false predictions. AlphaFold 3’s ability to predict protein-ligand interactions at near-atomic accuracy changes the economics and timeline of early-stage drug discovery.
David Baker and the De Novo Revolution
While AlphaFold addressed the prediction problem — given a sequence, what shape will it fold into — David Baker’s work addressed the design problem: given a desired shape or function, what sequence will fold into it?
Baker’s lab developed RFdiffusion, a deep-learning framework similar to those used for AI image generation, that enables users to create completely novel proteins based on molecular specifications — de novo designs not derived from any natural protein. The implications are profound: rather than screening through libraries of natural or slightly modified proteins to find one that performs a desired function, researchers can specify the function they want and computationally generate protein sequences predicted to achieve it.
In December 2025, Baker’s lab released RFdiffusion3 — described as their most powerful and versatile protein engineering technology to date, capable of designing proteins that interact with any type of molecule including DNA, enabling applications in gene therapy and synthetic biology that earlier versions could not address. Simultaneously, the lab published a Nature Methods paper on atom-level enzyme active site scaffolding using RFdiffusion2, demonstrating designed enzymes with catalytic efficiency nearly on par with those found in nature — a capability that enables rational design of industrial enzymes, biosensors, and therapeutic enzymes from scratch.
A September 2025 paper from Baker’s lab applied the design approach to one of medicine’s most challenging problems: controlling drugs once they are inside the body. The study designed AI-generated molecular switches for interleukin-2, a powerful immune cytokine used in cancer therapy that is notorious for toxic side effects. The AI-designed protein switch applied to IL-2 allowed activated human immune cells to be silenced on demand upon addition of an effector molecule — providing an off-switch for a cancer immunotherapy drug that previously had no off-switch. This capability — designing proteins that respond to external signals to modulate drug activity — represents a level of control over therapeutic biologics that has not previously been achievable.
The Clinical Translation Moment
Isomorphic Labs, the drug discovery subsidiary of Alphabet spun out of Google DeepMind, is the company closest to bringing AI-designed drugs to patients. Founded in 2021 specifically to apply AlphaFold’s structural biology insights to pharmaceutical development, Isomorphic raised $600 million in a Series A in March 2025, led by Thrive Capital. The company has 17 active drug development programs across oncology, immunology, and cardiovascular disease, with the first AI-designed cancer drug set to enter Phase 1 clinical trials by end of 2026.
Isomorphic Labs president Colin Murdoch confirmed in an interview with Fortune that four years after its founding, the company is preparing to dose the first patients in clinical trials, with staffing underway ahead of the trials. At the January 2026 World Economic Forum, Demis Hassabis acknowledged that the first clinical trials were expected by end of 2026 — a slight delay from the 2025 timeline previously communicated, but still representing an imminent crossing of the line from computational design to human testing.
The significance of this milestone deserves emphasis. Drug discovery conventionally proceeds from target identification to hit discovery to lead optimization — a process that typically takes three to seven years before a compound enters clinical testing. Isomorphic’s claim, if validated in clinical results, is that AI-guided structural biology can compress this timeline substantially by eliminating the trial-and-error that characterizes conventional drug discovery. Isomorphic’s AI drug design engine achieves extremely precise predictions of protein-drug interactions and antibody structures, with some experts comparing it to a new baseline in the field.
The Open Science Dimension
A defining tension in the AI protein design field is the split between open-source tools and proprietary commercial platforms. AlphaFold 2 was released open-source, and its database of 214 million predicted structures is freely available to any researcher in the world. This openness accelerated a wave of academic drug discovery research — enabling laboratories that could never have conducted structural biology experimentally to pursue structure-informed drug design.
AlphaFold 3 broke with this model: when DeepMind released it, the model weights were withheld for six months, and when released, they came with a noncommercial license. Isomorphic’s IsoDDE proprietary drug design engine, built on AlphaFold 3’s capabilities, is entirely proprietary. More than 1,000 scientists signed an open letter protesting the restricted access to AlphaFold 3’s weights.
The competing open-source ecosystem has responded. Chai Discovery, backed by $225 million including participation from OpenAI, achieved a 77 percent success rate on the PoseBusters benchmark — slightly exceeding AlphaFold 3’s 76 percent performance. Boltz-1, from MIT and ETH Zurich, provides open-source biomolecular interaction modeling competitive with AlphaFold 3. ESMFold from Meta AI runs an order of magnitude faster without requiring multiple sequence alignments. RFdiffusion3 from Baker’s lab is open source. The open and proprietary ecosystems are racing each other, which is producing faster progress across both.
What Clinical Translation Actually Requires
The path from a computationally designed protein to an approved drug requires solving problems that structure prediction does not address. A protein designed to bind a target with high affinity in a computational model must be synthesized, expressed in a biological system, purified, characterized for stability and immunogenicity, tested in cells, tested in animals, and then tested in humans — each stage eliminating candidates that looked promising on the previous one.
The failure rate in drug development is not primarily a problem of not knowing what targets to hit or what molecules to try. It is a problem of compounds that work in computational models failing in cells, compounds that work in cells failing in animals, and compounds that work in animals failing in humans. AI-guided protein design addresses the early stages of this cascade — identifying better candidates to advance — but it cannot eliminate the biological complexity that causes failures at later stages.
De novo protein design allows for rapid, cost-effective, and customizable protein biologics, but clinical translational potential remains in early stages with limited discussion of translation from computational design to clinical validation. The first AI-designed drugs entering Phase 1 trials in 2026 will generate the most important data in the field’s history — not about computational capability, but about whether computationally designed proteins actually work as therapeutics in human patients.
Why It Matters
The bottleneck in pharmaceutical development has not been a shortage of disease targets or a shortage of chemical creativity. It has been a shortage of structural information — not knowing precisely how drug candidates would interact with their protein targets — and a consequent dependence on expensive, slow empirical screening of vast compound libraries. AlphaFold’s resolution of the structure prediction problem and Baker’s extension of it into de novo design represent the removal of that bottleneck. Whether the consequence is a two-fold or a ten-fold acceleration in drug discovery — and whether the first AI-designed drugs to enter clinical trials prove efficacious — will determine how transformative this technological moment actually is.
Closing Human Dimension
For sixty years, the proteins that biology has produced over three billion years of evolution set the ceiling on what drugs could do — because drugs worked by fitting into natural protein shapes, and natural protein shapes were determined by evolutionary constraints that had nothing to do with therapeutic utility. AI protein design removes that ceiling. A protein designed specifically to fit a disease-causing molecule, optimized from the first computational step for therapeutic function rather than evolutionary fitness, represents a qualitatively different kind of medicine. Whether the first generation of AI-designed proteins entering clinical trials in 2026 confirms that potential or reveals new obstacles, the field is no longer asking whether this approach is theoretically possible. It is asking whether it works in patients.
Sources
1. IntuitionLabs. “Isomorphic Labs & AlphaFold: AI Drug Discovery in Trials.” June 2026. https://intuitionlabs.ai/articles/isomorphic-labs-alphafold-ai-drug-discovery-trials — documents Isomorphic’s clinical trial timeline and WEF 2026 statements.
2. DeepCeutix. “Google Just Built the Best Drug Design AI in History. You Can’t Use It.” March 2026. https://deepceutix.com/insights/proprietary-ai-drug-design — documents 17 programs, Phase 1 by end 2026, Chai Discovery benchmark.
3. Clinical Trials Arena. “Isomorphic Labs prepares to launch trials for AI-designed drugs.” July 2025. https://www.clinicaltrialsarena.com/news/isomorphic-labs-prepares-trials-ai-designed-drugs/
4. EurekAlert / Precision Clinical Medicine. “AlphaFold 3 ushers in a new era for biomedical research and drug discovery.” October 2025. https://www.eurekalert.org/news-releases/1101499
5. PMC. “De novo protein design: a transformative frontier in clinical protein applications.” January 2026. https://pmc.ncbi.nlm.nih.gov/articles/PMC12958671/
6. Genetic Engineering & Biotechnology News. “RFdiffusion3 Now Open Source, Designs DNA Binders and Advanced Enzymes.” December 2025. https://www.genengnews.com/topics/artificial-intelligence/rfdiffusion3-now-open-source-designs-dna-binders-and-advanced-enzymes/
7. Genetic Engineering & Biotechnology News. “AI-Designed Protein Switches Control Drugs with Speed.” September 2025. https://www.genengnews.com/topics/artificial-intelligence/ai-designed-protein-switches-control-drugs-with-speed/
8. GeekWire. “UW Nobel winner’s lab releases most powerful protein design tool yet.” December 2025. https://www.geekwire.com/2025/uw-nobel-winners-lab-releases-most-powerful-protein-design-tool-yet/
9. PMC. “AlphaFold3: An Overview of Applications and Performance Insights.” April 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC12027460/
10. Angewandte Chemie. “Structure Prediction and Computational Protein Design for Efficient Biocatalysts and Bioactive Proteins.” December 2024. https://onlinelibrary.wiley.com/doi/10.1002/anie.202421686
Idea originated at artificialideas.org. Article researched and written by Claude Sonnet 4.6. Published at artificialideas.org.
Artificial Ideas 