Five years ago, a quiet revolution began to unfold in the world of biology. Not with a bang, but with a complex algorithm named AlphaFold 2. Developed by Google DeepMind, it accomplished something scientists had been chasing for half a century: accurately predicting the three-dimensional structures of proteins, the very workhorses of life, with atomic precision. It was an achievement so profound that in 2024, its co-leader, John Jumper, alongside DeepMind CEO Demis Hassabis, was awarded a Nobel Prize in Chemistry.
For those of us watching from the sidelines, it felt like science fiction becoming reality. The “protein folding problem” had long been a grand challenge, a seemingly insurmountable hurdle. Now, with AlphaFold, results that once took months of painstaking lab work could be generated in mere hours. The hype, as you might imagine, was immense. But now that the initial fervor has settled, what’s the true legacy of AlphaFold? How are researchers leveraging this powerful tool, and what lies ahead for this game-changing AI? I recently had the privilege of speaking with John Jumper himself, along with other scientists, to find out.
The AlphaFold Revolution: Five Years On
“It’s been an extraordinary five years,” Jumper reflects, a chuckle in his voice, “It’s hard to remember a time before I knew tremendous numbers of journalists.” His journey to this point is a fascinating one. Fresh from a theoretical chemistry PhD in 2017, Jumper joined DeepMind on a secret project – one that aimed to predict protein structures. Just three years later, AlphaFold 2 made its stunning debut, and the rest, as they say, is history.
Proteins are fundamental to every living thing. They build our tissues, transport oxygen, power our immune systems, and enable every biological process imaginable. But to truly understand their function, you need to know their shape. Proteins are long strings of amino acids that fold into intricate 3D knots, and predicting that final shape from the initial string is incredibly difficult – an untwisted string offers few clues, and the possibilities are astronomical.
Jumper and his team tackled this problem using a neural network called a transformer, the same foundational technology behind today’s large language models. A key to their success? Rapid prototyping. “We got a system that would give wrong answers at incredible speed,” Jumper explains, highlighting how this allowed them to quickly test adventurous new ideas. They fed the network vast amounts of protein structure data, including evolutionary insights, and the system exceeded all expectations. “We were sure we had made a breakthrough,” he says.
Since its initial release, AlphaFold has continued to evolve. AlphaFold Multimer expanded its capabilities to predict structures involving multiple proteins, and AlphaFold 3 arrived as the fastest iteration yet. DeepMind also unleashed AlphaFold on UniProt, a global protein database, ultimately predicting the structures of some 200 million proteins – virtually every known protein to science. Despite this monumental success, Jumper remains grounded. “That doesn’t mean that we’re certain of everything in there,” he cautions. “It’s a database of predictions, and it comes with all the caveats of predictions.”
Beyond the Lab Bench: “Off-Label” Successes
What truly surprised Jumper wasn’t just AlphaFold’s accuracy, but how quickly and creatively scientists adopted it. He expected impact, but perhaps a few iterations down the line. Instead, researchers immediately began applying it in myriad ways. “I’ve been shocked at how responsibly scientists have used it,” he notes, “neither too much nor too little.”
He recalls a research group using AlphaFold to study disease resistance in honeybees. “I never would have said, ‘You know, of course AlphaFold will be used for honeybee science,’” he laughs. These are what he calls “off-label” uses, where AlphaFold’s predictive power has opened up entirely new research techniques. One prominent example is in protein design. David Baker, a computational biologist and co-winner of last year’s Nobel with Jumper and Hassabis, has been a pioneer in creating synthetic proteins for specific tasks, like treating diseases or breaking down plastics.
Baker’s team, among others, has developed their own tools, like RoseTTAFold, but also integrates AlphaFold Multimer into their workflow. Essentially, if AlphaFold confidently predicts the structure of a designed protein, they proceed; if it’s unsure, they don’t. “That alone was an enormous improvement,” Jumper says, estimating it makes the design process ten times faster.
Another fascinating “off-label” use transforms AlphaFold into a kind of biological search engine. Jumper highlights two groups studying human sperm-egg fertilization. Knowing one key egg protein, they ran all 2,000 human sperm surface proteins through AlphaFold, seeking a confident interaction. They found one, which they later confirmed in the lab. “You would never do 2,000 structures looking for one answer” conventionally, Jumper explains. “This kind of thing I think is really extraordinary.”
Real-World Impact and the Road Ahead
To gauge AlphaFold’s long-term impact, I circled back with Kliment Verba, a molecular biologist at the University of California, San Francisco, whom I’d spoken with at AlphaFold 2’s debut. Five years on, his verdict is clear: “It’s an incredibly useful technology, there’s no question about it. We use it every day, all the time.”
However, Verba is also candid about its limitations. While revolutionary, AlphaFold isn’t perfect, particularly when predicting interactions between multiple proteins or between proteins and smaller molecules – critical aspects for pathogen study or drug development. “There are many cases where you get a prediction and you have to kind of scratch your head,” he says. “Is this real or is this not? It’s not entirely clear—it’s sort of borderline.” He draws a parallel to generative AI: “It’s sort of the same thing as ChatGPT. You know—it will bullshit you with the same confidence as it would give a true answer.”
Despite these caveats, Verba’s team uses AlphaFold 2 and 3 (leveraging their distinct strengths) to run virtual experiments, narrowing down their focus or even deciding against certain lab experiments altogether. “It hasn’t really replaced any experiments, but it’s augmented them quite a bit,” he notes, emphasizing the significant time savings.
A New Wave of Innovation: Pushing the Boundaries
AlphaFold’s success has spawned a new wave of innovation. Startups and university labs are now building specialized tools atop its foundation, particularly for drug discovery. This year, MIT researchers collaborated with AI drug company Recursion to create Boltz-2, which predicts not just protein structures, but also how well potential drug molecules will bind to their targets.
Just last month, Genesis Molecular AI released Pearl, another structure prediction model. The company claims Pearl offers greater accuracy than AlphaFold 3 for specific queries relevant to drug development, and its interactive nature allows developers to feed in additional data to refine predictions. Evan Feinberg, Genesis Molecular AI’s CEO, acknowledges AlphaFold’s leap but insists there’s still much to do. “We’re still fundamentally innovating, just with a better starting point than before.”
Their focus? Pushing the margin of error below AlphaFold’s de facto industry standard of less than two angstroms, aiming for under one angstrom – the width of a single hydrogen atom. Michael LeVine, VP of modeling and simulation at Genesis Molecular AI, explains why this precision is crucial: “Small errors can be catastrophic for predicting how well a drug will actually bind to its target.” Chemical forces that interact at one angstrom might cease to do so at two, transforming a “they will never interact” into a “they will.”
Given this flurry of activity, how soon can we expect new drugs on the market? Jumper is pragmatic. Protein structure prediction is just one piece of a very complex puzzle. “This was not the only problem in biology. It’s not like we were one protein structure away from curing any diseases.” He contextualizes it: if finding a protein’s structure cost $100,000 in the lab, “If we were only a hundred thousand dollars away from doing a thing, it would already be done.”
At the same time, Jumper recognizes the power they’ve unleashed. Researchers are actively looking for ways to maximize this technology’s impact. “We’re trying to figure out how to make structure prediction an even bigger part of the problem, because we have a nice big hammer to hit it with.” In essence, he adds, they’re looking to turn everything into a nail. “How do we make this thing that we made a million times faster a bigger part of our process?”
Fusing AI Power: What’s Next for Jumper and AlphaFold?
So, what’s next for the Nobel laureate himself? Jumper envisions fusing AlphaFold’s deep, narrow predictive power with the broad capabilities of large language models (LLMs). “We have machines that can read science. They can do some scientific reasoning,” he explains. “And we can build amazing, superhuman systems for protein structure prediction. How do you get these two technologies to work together?”
His vision brings to mind AlphaEvolve, another DeepMind project that uses an LLM to generate solutions and a second model to filter them. While Jumper keeps details close to his chest, he’s confident about the direction. “I won’t say too much on methods, but I’ll be shocked if we don’t see more and more LLM impact on science,” he muses. “I think that’s the exciting open question that I’ll say almost nothing about. This is all speculation, of course.”
A Nobel Laureate’s Next Chapter
At 39, John Jumper is the youngest chemistry laureate in 75 years, a fact that gives him pause. “It worries me,” he admits. “I’m at the midpoint of my career, roughly.” His strategy for what comes next is refreshingly humble and wise. “I guess my approach to this is to try to do smaller things, little ideas that you keep pulling on. The next thing I announce doesn’t have to be, you know, my second shot at a Nobel. I think that’s the trap.”
The Unfolding Future
AlphaFold’s debut marked a pivotal moment, transforming how we approach biological research and accelerating discovery in ways previously unimaginable. Its journey from a secret project to a Nobel-winning tool, used daily by countless scientists, underscores the profound impact of AI when applied to fundamental scientific challenges. While the path to new drugs remains long and complex, AlphaFold has undeniably equipped researchers with a “big hammer,” empowering them to ask and answer questions that were once out of reach.
The future, as Jumper suggests, lies in an even deeper integration of AI — combining the precision of structure prediction with the expansive reasoning of LLMs. It’s a testament not just to technological ingenuity, but to the relentless curiosity and collaborative spirit of the scientific community. The next five years will undoubtedly bring further breakthroughs, demonstrating that AlphaFold wasn’t the final answer, but rather a spectacular opening chapter in a story that continues to unfold.
