In 2017, theoretical chemist John Jumper took a gamble on a rumor. Fresh from finishing his PhD, he’d heard whispers that Google DeepMind was shifting from game-playing AIs to a clandestine effort to predict the structures of proteins. He applied, joined the team, and within three years became co-lead of one of the most consequential scientific breakthroughs in decades.
That breakthrough was AlphaFold 2, the AI system unveiled five years ago this week and capable of predicting protein structures with atom-level accuracy—returning results in hours instead of the months required by traditional lab techniques. Alongside CEO Demis Hassabis, Jumper delivered what many in biology had assumed would take generations. The achievement earned the pair the 2024 Nobel Prize in chemistry, celebrated not just as an advance in computing but as the solution to a grand challenge that had lingered for half a century.
In interviews with MIT Technology Review, Jumper reflects on the half-decade since AlphaFold’s debut. The intensity still surprises him. “It’s hard to remember a time before I knew tremendous numbers of journalists,” he jokes.
After the initial system came AlphaFold Multimer, capable of handling structures composed of multiple proteins, and later AlphaFold 3, the fastest and most refined version to date. DeepMind also deployed the system across UniProt, a massive global protein database. Today, AlphaFold has predicted around 200 million protein structures—essentially every protein currently known to science. Yet Jumper stresses the outputs remain predictions. “It comes with all the caveats of predictions,” he says, noting that the database should be interpreted with appropriate skepticism.
Protein structure determination remains one of biology’s most stubborn puzzles. Proteins, made of amino acid chains that fold into intricate forms, can theoretically assume astronomical numbers of shapes. Predicting the one correct structure for each protein was long considered intractable. Jumper’s team built AlphaFold 2 on transformers—the same type of neural network behind large language models—and pushed forward by developing prototypes that could fail quickly and inform rapid iteration. “We got a system that would give wrong answers at incredible speed,” he recalls. “That made it easy to start becoming adventurous with the ideas you try.”
What the team didn’t foresee was how fast the global research community would adopt it. From the beginning, scientists downloaded the software and used it across far-flung disciplines. One group investigated disease resistance in honeybees. “I never would have said, ‘Of course AlphaFold will be used for honeybee science,’” says Jumper. Other teams repurposed AlphaFold in unexpected ways, including using it as a kind of biological search engine. One pair of labs ran all 2,000 human sperm surface proteins against a known egg protein, using AlphaFold to find a likely binding partner—later verified experimentally.
Researchers in protein design have been among the most enthusiastic adopters. David Baker’s lab at the University of Washington—whose work in synthetic proteins earned Baker a share of last year’s Nobel Prize—now uses AlphaFold Multimer to assess whether designed proteins are likely to fold as intended. If AlphaFold predicts a design confidently, they synthesize it. If not, they discard it. Jumper estimates this approach accelerates design cycles tenfold.
Still, AlphaFold has limitations. Kliment Verba, a molecular biologist at UC San Francisco who spoke with MIT Technology Review both at AlphaFold’s release and now, says the tool is indispensable but imperfect. Predicting interactions between multiple proteins—or between proteins and smaller molecules—remains harder, and many results sit in a gray zone of uncertainty. “There are many cases where you get a prediction and you have to scratch your head,” Verba says. “Is this real or is this not?” He compares it to large language models: “It will bullshit you with the same confidence as it would give a true answer.”
To address AlphaFold’s gaps, a new ecosystem of startups and academic labs has emerged. A collaboration between MIT researchers and Recursion produced Boltz-2, which predicts not just protein structures but how strongly potential drugs will bind to specific targets. The startup Genesis Molecular AI recently released Pearl, a model it claims outperforms AlphaFold 3 on certain drug-relevant problems. These next-gen tools aim to push structural prediction accuracy from under two angstroms—AlphaFold’s benchmark—to under one angstrom, the width of a hydrogen atom. Small differences at this scale can determine a drug’s ability to bind or fail.
Yet even with these advances, Jumper warns against expecting immediate medical revolutions. Protein structure prediction is just one piece of the drug-discovery pipeline. “It’s not like we were one protein structure away from curing any diseases,” he says. But he notes that researchers are eager to expand how far structural prediction can take them. “We have a nice big hammer,” he says. “How do we make this thing that we made a million times faster a bigger part of our process?”
As for the future, Jumper is now exploring how to merge specialized systems like AlphaFold with broad-reasoning large language models. He envisions hybrids that can read scientific literature, reason about experiments, and combine that knowledge with ultra-accurate molecular prediction. MIT Technology Review points to DeepMind’s AlphaEvolve—an LLM-guided system that generates and filters solutions in math and computer science—as an example of what such fusion might look like.
At 39, Jumper became one of the youngest chemistry laureates in modern history, a fact that he admits makes him uneasy. “I guess my approach is to try to do smaller things, little ideas that you keep pulling on,” he says. His next breakthrough, he insists, doesn’t need to be another Nobel-scale event.
Five years on, AlphaFold stands as both a triumph and a beginning. It solved a problem that stumped generations, but in doing so opened a larger frontier—a new era in which AI and biology coevolve, and where researchers, armed with tools once considered impossible, must decide how best to wield them.
