In 2017, shortly after earning his PhD in theoretical chemistry, John Jumper caught wind of speculation that Google DeepMind had transitioned from creating AI for games with superhuman abilities and was embarking on a clandestine endeavor to forecast protein structures. He submitted an application for employment.
Merely three years later, Jumper marked a remarkable victory that took many by surprise. Alongside CEO Demis Hassabis, he had co-directed the creation of an AI solution named AlphaFold 2, which could predict protein structures to within atomic dimensions, matching the precision of labor-intensive lab techniques and achieving this many times more rapidly—delivering outcomes in hours rather than months.
AlphaFold 2 had addressed a monumental challenge in biology that had persisted for 50 years. “This is the reason I initiated DeepMind,” Hassabis remarked to me a few years back. “In fact, it’s the driving force behind my entire career in AI.” In 2024, Jumper and Hassabis were awarded a Nobel Prize in chemistry.
This week marks five years since AlphaFold 2’s introduction took the scientific community by surprise. Now that the excitement has subsided, what real impact has AlphaFold had? How are researchers utilizing it? What does the future hold? I spoke with Jumper (and several other scientists) to learn more.
“It’s been an extraordinary five years,” Jumper states, chuckling: “It’s difficult to recall a time before I knew a multitude of journalists.”
Following AlphaFold 2, AlphaFold Multimer was introduced, capable of predicting structures comprising more than one protein, and then AlphaFold 3, the quickest version to date. Google DeepMind also released AlphaFold for UniProt, an expansive protein database utilized and updated by millions of researchers globally. It has now predicted the structures of nearly 200 million proteins, encompassing almost all known to science.
Despite his success, Jumper remains humble regarding AlphaFold’s accomplishments. “That doesn’t imply we are certain about everything contained within,” he states. “It’s a repository of predictions, and it includes all the uncertainties associated with predictions.”
A challenging issue
Proteins serve as the biological engines that enable living organisms to function. They constitute muscles, horns, and feathers; they transport oxygen throughout the body and transmit signals between cells; they activate neurons, break down food, and energize the immune system; and much more. However, precisely understanding a protein’s function (and the potential role it may have in various illnesses or treatments) requires deciphering its structure—and that is quite complex.
Proteins are constructed from chains of amino acids that chemical forces entwine into intricate knots. An uncoiled chain offers minimal insight into the configuration it will adopt. Theoretically, most proteins could assume an inconceivable number of potential shapes. The challenge is to identify the correct one.
Jumper and his team developed AlphaFold 2 utilizing a type of neural network known as a transformer, the same technology that supports large language models. Transformers excel at focusing on specific segments of a larger conundrum.
However, Jumper attributes much of the success to rapidly creating a prototype model they could test efficiently. “We established a system that could generate incorrect answers at remarkable speed,” he explains. “This facilitated a willingness to explore bold ideas.”
They filled the neural network with comprehensive information about protein structures, including how proteins from various species exhibit similar shapes over time. It performed even better than anticipated. “We were confident we had made a significant leap,” Jumper states. “We were convinced this was a remarkable advance in concepts.”
What he had not anticipated was that scientists would immediately download his software and apply it for countless diverse purposes. Typically, it’s the later iterations that make the most substantial impact after resolving initial issues, he notes: “I’ve been astonished by the responsible manner in which researchers have utilized it, in terms of interpreting it, and employing it with the level of trust it warrants—neither excessive nor insufficient.”
Any particular projects stand out?
Honeybee research
Jumper highlights a research team using AlphaFold to investigate disease resistance in honeybees. “They sought to comprehend a specific protein in their studies related to colony decline,” he remarks. “I never would have predicted, ‘Of course, AlphaFold will be utilized for honeybee research.’”
He also points out several instances of what he refers to as off-label applications of AlphaFold, “in the sense that efficacy was not assured”—where the capacity to predict protein structures has created innovative research methodologies. “The first is clearly the advancements in protein engineering,” he specifies. “David Baker and others have certainly embraced this technology.”
Baker, a computational biologist at the University of Washington, was a co-recipient of last year’s Nobel Prize in chemistry, alongside Jumper and Hassabis, for his contributions to fabricating synthetic proteins designed to execute specific functions—such as treating health issues or decomposing plastics—more effectively than natural proteins can.
Baker and his team have created their own tool based on AlphaFold, named RoseTTAFold. They have also experimented with AlphaFold Multimer to forecast which of their synthetic protein designs would be effective.
“Essentially, if AlphaFold confidently endorses the structure you aimed to design and then you execute it, and if AlphaFold expresses uncertainty, you refrain from proceeding. That alone marked a significant enhancement.” It can accelerate the design workflow by up to 10 times, Jumper states.
Another unconventional application that Jumper emphasizes: Transforming AlphaFold into a type of search engine. He cites two separate research teams examining how human sperm cells interact with eggs during fertilization. They were aware of one of the proteins involved but not the other, he says: “So, they took a known egg protein and tested all 2,000 human sperm surface proteins, discovering one that AlphaFold strongly indicated bound to the egg.” They were subsequently able to validate this in the laboratory.
“This idea that you can leverage AlphaFold for tasks previously impossible—you wouldn’t normally analyze 2,000 structures seeking a single answer,” he explains. “This kind of thing I think is truly remarkable.”
Five years later
When AlphaFold 2 was launched, I consulted a group of early users for their impressions. The feedback was positive, but the technology was too nascent to ascertain its long-term implications. I reconnected with one of those initial users to gather his thoughts five years later.
Kliment Verba, a molecular biologist managing a lab at the University of California, San Francisco, states, “It’s an incredibly beneficial technology, there’s no doubt about it.” He adds, “We rely on it daily, nonstop.”
However, it’s far from flawless. Numerous scientists employ AlphaFold to explore pathogens or devise medications. This involves examining interactions among multiple proteins or between proteins and even smaller molecules within the body. Nevertheless, AlphaFold is recognized for having reduced accuracy when predicting interactions involving multiple proteins over time.
Verba mentions that he and his colleagues have utilized AlphaFold long enough to understand its limitations. “There are various situations where you receive a prediction that leaves you puzzled,” he indicates. “Is this genuine or not? It’s somewhat ambiguous—sort of borderline.”
“It’s somewhat akin to ChatGPT,” he continues. “You know—it can mislead you with the same confidence as it would provide a truthful answer.”
Nonetheless, Verba’s team employs AlphaFold (both versions 2 and 3, as they possess distinct strengths, he asserts) to conduct virtual simulations of their experiments before executing them in the lab. Utilizing AlphaFold’s findings, they can refine the scope of an experiment—or determine that it’s not feasible.
It can significantly conserve time, he notes: “It hasn’t replaced any experiments, but it has greatly complemented them.”
New era
AlphaFold was intended to serve various functions. Presently, numerous startups and academic labs are capitalizing on its achievements to create a fresh wave of tools tailored for drug discovery. This year, a collaboration between MIT researchers and the AI pharmaceutical company Recursion yielded a model named Boltz-2, which predicts not just protein structures but also the binding efficacy of potential drug compounds to their targets.
Last month, the startup Genesis Molecular AI unveiled another structure prediction model called Pearl, which the company claims is more precise than AlphaFold 3 for specific inquiries crucial for drug development. Pearl is interactive, enabling drug developers to input any supplementary data they may possess to guide its forecasts.
AlphaFold represented a significant advancement, yet there is more to accomplish, asserts Evan Feinberg, CEO of Genesis Molecular AI: “We are still fundamentally innovating, but with a better foundation than before.”
Genesis Molecular AI is striving to decrease margins of error from under two angstroms, the standard set by AlphaFold, to below one angstrom—one ten-millionth of a millimeter, or the diameter of a single hydrogen atom.
“Minor inaccuracies can have severe consequences for predicting how effectively a drug will actually bind to its target,” explains Michael LeVine, vice president of modeling and simulation at the company. This is due to the fact that chemical forces that engage at one angstrom may cease to do so at two. “It can shift from ‘They will never interact’ to ‘They will,’” he states.
Given the immense activity in this domain, how soon can we anticipate new types of drugs to enter the market? Jumper takes a realistic perspective. Protein structure forecasting is merely one of many steps, he remarks: “This was not the sole issue in biology. We were not just one protein structure away from curing all diseases.”
Consider it this way, he suggests. Discovering a protein’s structure may have previously incurred a $100,000 cost in the lab: “If we were merely a hundred thousand dollars from achieving something, it would have already been accomplished.”
Meanwhile, researchers are seeking ways to maximize this technology’s potential, according to Jumper: “We aim to figure out how to make structure prediction a more substantial component of the problem, as we have a powerful tool to address it.”
In other words, do they want to turn everything into nails? “Yes, let’s convert things into nails,” he responds. “How can we integrate this method that we’ve accelerated millions of times into a larger aspect of our approach?”
Upcoming developments
What’s Jumper’s next goal? He aspires to merge the deep yet narrow capabilities of AlphaFold with the expansive scope of large language models.
“We possess machines capable of digesting scientific literature. They can engage in some scientific reasoning,” he notes. “And we can build extraordinary, superhuman systems for protein structure prediction. How can we get these two technologies to collaborate?”
This reminds me of a system named AlphaEvolve, which another team at Google DeepMind is constructing. AlphaEvolve employs an LLM to generate potential solutions to a problem and a second model to assess them, filtering out the unviable options. Researchers have already utilized AlphaEvolve to achieve several meaningful discoveries in mathematics and computer science.
Is this what Jumper envisions? “I won’t elaborate too much on our methods, but I would be surprised if we don’t witness increasingly significant LLM influence on science,” he observes. “I believe that’s the intriguing open question about which I will say little. This is all conjecture, of course.”
At 39, Jumper became a Nobel Prize laureate. What lies ahead for him?
“It does concern me,” he remarks. “I believe I might be the youngest chemistry laureate in 75 years.”
He continues: “I’m at the midpoint of my career, roughly. My approach is to focus on smaller projects, gradual ideas that you continually delve into. The next announcement I make doesn’t necessarily need to be, you know, my second attempt at a Nobel Prize. I think that’s a trap.”
