In reality, in October 2024, the Nobel Prize in Chemistry went to software, in a sense. Half the honor recognized AlphaFold2, a model that predicted the structure of nearly all known proteins. The other half recognized something harder: teaching computers to invent proteins that never existed in nature. In reality, Scientists described the shift as biology learning to write, not just read. Few awards have ever marked such a clean changing of the guard. Chemistry will not feel the same again. That is the story of AI protein design, told here.
In reality, this guide unpacks AI protein design and the prize behind it. You will meet the three winners and the models that earned the honor. Also, we will keep the chemistry readable and the claims honest. Also worth knowing: the science is already producing vaccine parts and medicine candidates. Each section below stands on its own, so skimming works too.
In this guide
- AI protein design: the 30-second answer
- What you’ll learn
- Reading proteins: the prediction half of the prize
- Writing proteins: David Baker and RFdiffusion
- AlphaProteo: DeepMind answers with its own designer
- From designed scaffolds to real medicines
- What AI protein design still cannot do
- Frequently asked questions
- The bottom line
AI protein design: the 30-second answer
In short, here is the short version. AI protein design teaches software to invent new proteins with useful shapes and jobs. In short, Prediction, by contrast, reads proteins that already exist. In 2024, both skills shared the Nobel Prize in Chemistry. Designed proteins now reach vaccine scaffolds, enzymes and therapeutic candidates. Overall, the field moved from reading nature to writing it. Both skills chase one goal: better medicines, faster.
Key takeaways
- Meanwhile, the Nobel Prize Chemistry 2024 honored computational protein prediction and design.
- Demis Hassabis and John Jumper of Google DeepMind won for AlphaFold2.
- David Baker of the University of Washington won for computational protein design.
- His lab’s RFdiffusion generates brand new proteins, including protein binders for chosen targets.
- DeepMind’s AlphaProteo designs binders too, and the company reports high experimental success rates.
What you’ll learn
The route through this guide
- Why reading proteins earned a Nobel before writing them did
- Meanwhile, how AlphaFold2 changed prediction almost overnight
- What RFdiffusion and AlphaProteo actually build
- Where designed proteins are already heading, from vaccines to antibodies
- What this science still cannot do
- Where to read deeper inside this series
Reading proteins: the prediction half of the prize
In particular, Proteins are the machinery of life, and their shapes decide their jobs. For half a century, predicting those shapes from amino acid sequences was biology’s unsolved grand challenge. In particular, AlphaFold2 cracked it, predicting structures for nearly all known proteins. Demis Hassabis and John Jumper of Google DeepMind led the work and earned half of the 2024 Nobel. Medicine gained a working atlas of its favorite targets.
Why does prediction matter for medicines? Because drugs work by binding to proteins, like keys fitting locks. When a model can predict every lock, designing keys stops being guesswork. Also, our AlphaFold 3 explained walkthrough covers where the newest model goes next. AI protein design begins where prediction ends, with shapes nobody has measured.
Prediction, however, reads a world that already exists. Also, Design runs the other way and asks for proteins nobody has ever seen. That harder, opposite problem is what David Baker solved well enough to win. AI protein design, in short, completed what prediction started.
In fact, the atlas effect deserves a pause. Before prediction models, structural work often consumed years per protein. In fact, Now a student can request a structure and receive a ranked answer in minutes. Whole labs reorganized their workflows around that change. For drug hunters, the starting line moved closer to the finish. Speed here changed behavior, not just convenience.
In fact, Skeptics asked fair questions along the way, as they should. Models sometimes miss unusual proteins or rare modifications. In fact, Validation experiments kept the field honest through every claim. That culture of checking is part of why the prize felt earned. Trust here was built, not granted.
In fact, one more effect deserves mention. Prediction databases keep growing as science sequences more species. Every new structure enriches the training ground for design. Reading and writing, in other words, feed each other. The loop is the real story.
Writing proteins: David Baker and RFdiffusion
David Baker of the University of Washington spent decades on computational protein design. His group learned to invent folded proteins that never existed in nature. The Nobel committee honored exactly that achievement with the other half of the 2024 Chemistry prize. AI protein design earned the field’s highest scientific honor.
The lab’s flagship tool is RFdiffusion, a diffusion model built for proteins. Diffusion models learn from noisy data, then generate clean samples, the same trick behind modern image generators. Applied to biology, RFdiffusion generates brand new protein structures on demand. Among its outputs are protein binders, shapes built to grip chosen targets. Binders can sit on a disease protein and interfere with its work. For drug hunters, that means candidate shapes on demand rather than after years of modeling.
Why are protein binders such a big deal? A binder that grips a disease protein can change how it behaves. That makes binders a natural starting point for therapies and research tools. Binders are also the clearest proof that AI protein design genuinely works. A designed shape that folds and grips is design proven end to end.
Diffusion design also invites creativity. Researchers can request a shape first and ask what it might do later. That reverses the old order, where function came first and structure followed. In design mode, structure is a starting point instead of a conclusion. The habit feels strange at first, and researchers report adapting quickly.
Baker’s influence also runs through people. Former students now lead design groups across academia and industry. Ideas spread faster when a field shares its tools openly. That culture, colleagues say, is part of what the Nobel recognized. Science here looks like a relay, not a solo run.
A practical note closes this section. Design tools complement, rather than replace, the slower work of characterization. Every generated structure still meets spectroscopy, purification and functional assays. The wet lab remains the final judge of every design. Software simply arrives with far better guesses.
AlphaProteo: DeepMind answers with its own designer
DeepMind answered in September 2024 with AlphaProteo, its own design system for protein binders. The system designs binders for targets including VEGF-A, a protein tied to cancer and eye disease. Company reports described high experimental success rates across the targets it tested. Treat those numbers as early coverage, since independent replication takes time. Even so, two rival design engines now serve the field.
The rivalry matters for medicine, not just for bragging rights. Two strong design engines mean faster iteration for laboratories everywhere. Competition also keeps both teams honest about their limits. As a result, AI protein design tooling keeps getting richer. Meanwhile, the field’s shared bench keeps growing. Readers benefit from watching both engines improve.
Progress of this kind always raises a fair question. Who checks the designs, and how quickly can they reach patients? Those answers belong to laboratories, regulators and clinical trials. Software proposes, but biology and medicine dispose. That division of labor is a feature, not a delay.
Clinical trials will arbitrate the rivalry eventually. Designed binders must show real benefit in patients, not just elegant lab data. Until then, both systems compete on speed, accuracy and experimental success rates. That competition benefits every laboratory downstream. Patients rarely follow the race, yet they will feel the results.
VEGF-A deserves a closer look, because it shows the targeting logic. The protein drives blood vessel growth and is tied to cancer and eye disease. A binder designed to grip it becomes a research tool the moment it folds. Therapeutic uses would come later, after the usual evidence. Targets like this one are where design earns its keep.
What should a reader take from this race? Treat design systems as powerful collaborators with sharp edges. Watch for validation studies, not just announcement posts. Also, expect targets to grow more ambitious as confidence builds. The scoreboard here is ultimately patients served.
From designed scaffolds to real medicines
AI protein design already serves practical roles. Designed proteins work as vaccine scaffolds, as enzymes and as therapeutic candidates. AI-designed antibodies are moving toward and into clinical development. The distance from screen to shelf keeps shrinking. Patients, meanwhile, watch a new kind of medicine arrive.
Antibodies deserve special attention here. These immune proteins can be engineered to grip disease targets with high precision. Design tools aim them at targets that conventional screening struggles to reach. For patients, that means more candidates for targets once considered unreachable. It also means faster iteration between design and lab.
Vaccine scaffolds show the flexibility nicely. A designed scaffold presents pieces of a pathogen to the immune system in a stable frame. Enzymes designed this way join the list as well. Each use case needs the same core skill: shaping a protein for a job. That is precisely what the new tools deliver.
Clinical development is the gate that matters most. Designed candidates must clear the same evidence bar as any other medicine. Early wins will look small: a trial here, a cleared hurdle there. Over a decade, though, those small wins compound. That is usually how revolutions actually arrive. Evidence, not enthusiasm, moves candidates forward.
Thinking about how this connects to the rest of the cluster helps. Design needs readable biological instructions, which is where the Evo 2 DNA model enters. Small molecule work follows a parallel track covered in our generative AI molecule design guide. Together, these tools sketch an AI protein design pipeline with biology as the material. Each piece covers one stage of the same journey.
The protein design stack, from reading to writing
| Layer | What it does | Named example |
|---|---|---|
| Read existing proteins | Predicts structures for nearly all known proteins | AlphaFold2 |
| Generate new shapes | Diffusion model that invents brand new proteins | RFdiffusion |
| Design binders | Builds protein binders for chosen targets, including VEGF-A | AlphaProteo |
| Build therapeutics | Vaccine scaffolds, enzymes and therapeutic candidates | Designed antibodies entering clinical development |
| Prove it in the lab | Experiments confirm whether designs fold and work | Wet lab validation |
Reading that stack top to bottom shows the journey. Software reads, then generates, then targets, then builds, then proves. Each arrow between rows still needs human hands. That is what makes the stack a partnership rather than an assembly line. Overall, progress flows in both directions.
Regulators will move at their own careful pace. That slowness protects people, and impatient readers should expect it. Meanwhile, research applications multiply quickly because they carry lower risk. Tools for the lab arrive years before tools for the clinic. Both matter, and both are coming.
What AI protein design still cannot do
Every revolution has a fine print section, and this one deserves it. A designed protein that folds on screen may fail in a test tube. Experiments still decide everything, and they stay slow. Biology keeps a deep reserve of surprises for overconfident models. AI protein design inherits every one of those constraints.
The prize itself needs careful reading too. Nobel recognition honors a breakthrough, not a finished pipeline. Designed antibodies are moving toward and into clinical development, which is real but early. Company claims about success rates deserve the same polite skepticism as any press release. Independent replication is the currency this field trades in.
There is also a thinking style difference worth naming. Prediction runs on fast pattern recognition, while design demands slower, deliberate reasoning about purpose and structure. Our System One vs System Two AI explainer maps that contrast in detail. Both modes matter, and neither replaces the other. Good laboratories schedule time for both.
Where does the cornerstone fit? Start with our AI in medicine cornerstone to trace the whole journey from data to pharmacy. Protein design is one powerful chapter of that longer story. Read it when you want the map instead of the chapter. Also, the series keeps a dedicated reality check on limits.
None of these caveats cancels the prize. They simply mark where human judgment has to sit. A model can hand you a candidate worth testing, and only experiments can hand you a medicine. Keeping that order straight is the whole discipline. In addition, it protects budgets and patients alike. Patience, here, is a technical requirement.
Finally, a word on jobs. Protein design creates work for people who can question models and run careful experiments. Those skills sit close to existing lab training. Curiosity about machine learning helps, and so does patience. The field needs skeptics as much as enthusiasts.
One more marker of how fast this moved. In 2020, protein design papers were specialist reading. By 2024, the field had a Nobel Prize and commercial pipelines. By 2026, designed proteins were entering clinical development. Students watching that arc see why timing matters in science careers. The pace shows no sign of slowing yet.
Frequently asked questions
What did the Nobel Prize Chemistry 2024 honor?
It honored computational protein science. Demis Hassabis and John Jumper won for AlphaFold2, which predicted structures for nearly all known proteins. David Baker won for computational protein design, the art of inventing new proteins. The award split honored both reading and writing proteins.
What is AI protein design in simple terms?
It means using software to invent proteins with chosen shapes and functions. Scientists describe the goal as designing molecular machines on demand. Prediction reads what exists, while design invents what does not. Experiments then check whether each design works as intended.
What are protein binders?
They are designed proteins shaped to grip chosen targets. RFdiffusion and AlphaProteo both generate them. Binders can block a disease protein or mark it for study. Researchers reach for them first as tools, then as therapies. Most uses today are exploratory rather than clinical.
Who won the 2024 chemistry Nobel?
Demis Hassabis and John Jumper of Google DeepMind shared it with David Baker of the University of Washington. The first two built AlphaFold2 for prediction. Baker’s lab built RFdiffusion for design. Each half of the prize mapped one side of that divide.
Are AI-designed medicines actually being used?
Designed proteins already serve as vaccine scaffolds, enzymes and therapeutic candidates. AI-designed antibodies are moving toward and into clinical development. Routine clinical use is still ahead, not behind. Hospitals will see them gradually, as evidence accumulates.
Where should I start reading about AI in medicine?
Begin with our AI in medicine cornerstone for the full map. Then follow the AlphaFold 3 and Evo 2 articles for the technical engines. Each piece stands alone, so curiosity can lead. The cluster reads like one long story, told in chapters.
The bottom line
AI protein design turned biology from a science of reading into a craft of writing. The Nobel Prize Chemistry 2024 made that shift official. Machines now propose proteins that evolution never bothered to try. Laboratories turn those proposals into scaffolds, enzymes and therapeutic candidates. Patients should expect the first fruits quietly, trial by trial. Also, the toolkit keeps improving under competitive pressure.
Keep the celebration proportionate. Experiments still rule, designed antibodies are early, and biology loves a plot twist. Still, the direction is unmistakable: the next generation of medicines will be partly written, not only discovered. AI protein design earned its gold medal, and the experiments are just starting.
Sources
- The Nobel Prize in Chemistry 2024, popular information, nobelprize.org
- American Chemical Society coverage of the 2024 chemistry Nobel, October 2024, acs.org
- Google DeepMind, AlphaProteo announcement, September 2024, deepmind.google
- Institute for Protein Design publications on RFdiffusion, University of Washington
- Nature coverage of computational protein design and AlphaFold, nature.com
- Google DeepMind, AlphaFold2 and the protein structure database, deepmind.google