Imagine you’re a theoretical chemist, fresh out of your PhD program. You hear a whisper that Google DeepMind—the folks famous for building AIs that crush humans at complex games—is starting a secret project. The goal? To solve one of the biggest, gnarliest problems in biology.
So you apply for a job. And just three years later, you're on stage, sharing a Nobel Prize.
That’s basically the story of John Jumper. In 2020, he and DeepMind CEO Demis Hassabis co-led the team that created AlphaFold 2, an AI that did something scientists had been trying to do for half a century: predict how proteins fold into their complex 3D shapes. And it did it with terrifying accuracy, matching slow, expensive lab methods in a matter of hours, not months.
It’s been five years since AlphaFold first blew everyone’s minds. Now that the initial shock has worn off, I wanted to know what the real, on-the-ground impact has been. Is it actually being used? Did it live up to the hype? And what comes next? I got the chance to chat with Jumper himself to find out.
"It's been an extraordinary five years," he told me with a laugh. "It's hard to remember a time before I knew tremendous numbers of journalists."
The 50-Year-Old Puzzle That Stumped Scientists
So, why is this protein folding thing such a big deal anyway?
Think of proteins as the tiny biological machines that run pretty much everything in your body. They’re what make up your muscles, carry oxygen in your blood, fire your neurons, and fight off infections. They’re absolutely essential.
But here’s the tricky part. A protein starts as a simple string of amino acids. To do its job, that string has to twist and fold itself into a very specific, incredibly complex 3D knot. The problem is, for any given string, the number of possible shapes it could form is astronomically huge. Figuring out the one correct shape from that mess is the puzzle that stumped biologists for decades.
Jumper and his team cracked it using a type of AI called a transformer—the same fundamental tech that powers models like ChatGPT. Transformers are brilliant at seeing the big picture and understanding how different parts of a sequence relate to each other.
But Jumper credits a lot of their success to something simpler: building a system that could fail fast. "We got a system that would give wrong answers at incredible speed," he said. This let them be incredibly adventurous and test wild ideas without waiting weeks for results. They fed the AI a massive diet of known protein structures, and it worked even better than they dared to hope.
From Curing Disease to… Saving Bees?
When AlphaFold 2 was released, Jumper figured it would be a while before anyone really used it. Usually, it’s the third or fourth version of a new technology that has the real impact.
He was wrong. Scientists started downloading it and putting it to work immediately.
"I've been shocked at how responsibly scientists have used it," he admits, noting that they seem to have a good sense of when to trust its predictions and when to be skeptical.
I asked him if any particular projects stood out. He immediately brought up a group studying disease resistance in honeybees. "They wanted to understand this particular protein as they look at things like colony collapse," he says. "I never would have said, ‘You know, of course AlphaFold will be used for honeybee science.’”
That’s the beauty of a foundational tool—people find uses you never could have imagined. Jumper calls these "off-label uses," and he's fascinated by them.
One of the biggest has been in protein design. Scientists like David Baker at the University of Washington (who also shared the Nobel) are now creating brand new, synthetic proteins that don't exist in nature to do things like break down plastics or deliver drugs more effectively. AlphaFold has become a crucial reality check for them.
"Basically, if AlphaFold confidently agrees with the structure you were trying to design... you make it," Jumper explains. "And if AlphaFold says ‘I don’t know,’ you don’t make it." That simple step makes the design process about ten times faster.
Another wild use? Turning AlphaFold into a kind of biological search engine. Jumper told me about two different research teams trying to figure out how a human sperm cell latches onto an egg. They knew the protein on the egg, but not its partner on the sperm. So, they just ran all 2,000 human sperm proteins against the egg protein in AlphaFold. The AI confidently pointed to one that it was sure would stick. And it was right. That's something you could never do in a physical lab.
The Honeymoon is Over: What's It Really Like to Use AlphaFold?
This all sounds amazing, but what’s it like for the scientists in the trenches using it every day? To get a reality check, I caught up with Kliment Verba, a molecular biologist at UCSF. I’d spoken to him right after AlphaFold came out five years ago.
"It’s an incredibly useful technology, there’s no question about it," he told me. "We use it every day, all the time."
But it’s no magic bullet. Verba says they’ve learned its quirks and limitations. AlphaFold is great at predicting a single protein's structure, but it can get a bit shaky when predicting how multiple proteins interact with each other, which is crucial for 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."
He then made a comparison that I think anyone who's used AI recently will understand perfectly. "It’s sort of the same thing as ChatGPT," he said. "You know—it will bullshit you with the same confidence as it would give a true answer."
For Verba's team, AlphaFold hasn't replaced lab experiments, but it has supercharged them. They can run virtual experiments first to see what’s promising, saving huge amounts of time and resources on ideas that are likely to be dead ends.
The Next Wave Is Already Here
AlphaFold kicked the door open, and now a whole new wave of companies and labs is rushing through. Startups like Genesis Molecular AI and university collaborations are building tools specifically for drug discovery.
They’re building on AlphaFold’s success but aiming for even greater precision. Genesis’s model, Pearl, for example, is pushing for accuracy down to less than one angstrom—that’s the width of a single hydrogen atom. Why does that matter? As their VP Michael LeVine puts it, the chemical forces that make a drug bind to a protein can switch on or off between one and two angstroms. A tiny error can be the difference between a drug that works and one that does nothing.
But Jumper is pragmatic about how quickly this will lead to new medicines on the shelf.
"This was not the only problem in biology," he reminds me. "It’s not like we were one protein structure away from curing any diseases."
Think of it this way, he says: figuring out a protein’s structure used to cost maybe $100,000 in the lab. "If we were only a hundred thousand dollars away from doing a thing, it would already be done."
The real question now is how to best use this new superpower. The challenge has shifted. "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 other words, if all you have is a hammer, you start looking for ways to make everything look like a nail. "Yeah, let’s make things into nails," he says with a grin.
So, What's Next for the Guy Who Won a Nobel at 39?
It’s a little strange to think about your "next act" when your first one involved winning the most prestigious award in science before you even hit 40.
"It worries me," Jumper admits. "I believe I’m the youngest chemistry laureate in 75 years... I’m at the midpoint of my career, roughly."
So what does he want to do? He wants to bring two worlds of AI together: the deep, specialized intelligence of AlphaFold and the broad, general knowledge of large language models (LLMs).
"We have machines that can read science. They can do some scientific reasoning," he says. "And we can build amazing, superhuman systems for protein structure prediction. How do you get these two technologies to work together?"
He was cagey on the details, but it sounds like he envisions a future where AI can not only perform specific scientific tasks but also read the vast library of human knowledge, form hypotheses, and reason about science in a more holistic way.
It’s an incredible vision. But for now, he’s trying to avoid the pressure of finding his "second shot at a Nobel."
"I think that’s the trap," he says. Instead, his approach is to "try to do smaller things, little ideas that you keep pulling on." And in science, as in life, it's often those little threads that lead to the next big breakthrough.
