In December 2020, the protein structure prediction community gathered — virtually, as the pandemic raged — to hear the results of CASP14, the biennial competition that measures progress on one of biology's hardest problems. When AlphaFold 2's scores were revealed, the room went quiet. The system had achieved a median accuracy of 92.4 GDT, effectively matching experimental methods. The protein folding problem, a challenge that had resisted five decades of effort, had been solved. The scientist who led the team that built AlphaFold 2 was John Jumper — a physicist turned machine learning researcher who, at 35, had just delivered one of the most consequential scientific breakthroughs of the century.
A Physicist's Path to Biology
John Jumper's road to DeepMind wound through physics, not biology. He earned his undergraduate degree at Vanderbilt University, where he studied physics and mathematics, developing the kind of rigorous quantitative foundation that would later prove essential for attacking the protein folding problem.
He then moved to the University of Chicago for his doctoral work, joining the lab of Karl Freed, where he studied the physics of protein folding using molecular dynamics simulations. It was here that Jumper encountered the protein folding problem firsthand — and experienced its frustrations. Traditional physics-based simulations could model small proteins over short timescales, but the computational cost of simulating the folding of even a modest protein was staggering. The brute-force approach, while physically grounded, simply could not scale.
This experience gave Jumper a deep appreciation for both the physics of protein folding and the limitations of existing approaches. He understood the problem from first principles, but he also understood that a fundamentally different approach was needed.
Joining DeepMind
After completing his PhD, Jumper joined DeepMind in 2017, just as the company was beginning to explore the protein folding problem in earnest. DeepMind, founded by Demis Hassabis, had already demonstrated the power of deep learning and reinforcement learning through AlphaGo and other projects. Hassabis had identified protein structure prediction as a scientific problem where AI could make a transformative contribution, and he was assembling a team to tackle it.
Jumper was an unusual hire for DeepMind — a physicist with deep domain knowledge in protein biophysics, not a typical machine learning engineer. But that combination of skills turned out to be exactly what the project needed. The protein folding problem required not just powerful algorithms, but a deep understanding of the underlying science: the physics of atomic interactions, the constraints imposed by evolution, and the geometric rules that govern how amino acid chains pack into three-dimensional structures.
AlphaFold 1 and the Lessons of CASP13
The AlphaFold team's first major test came at CASP13 in 2018. AlphaFold 1 used deep neural networks to predict distances between pairs of amino acids, then used gradient descent to find 3D structures consistent with those predictions. The system won the competition convincingly, outperforming all other entries.
But Jumper and the team knew that AlphaFold 1 had significant limitations. Its two-stage pipeline — first predict distances, then optimize a structure — was cumbersome and prone to error propagation. The system struggled with larger proteins and produced structures that, while better than the competition, were not accurate enough to be truly useful for biology. The team needed a fundamentally different architecture.
Building AlphaFold 2: The Evoformer
Over the next two years, Jumper led the design and development of AlphaFold 2, a ground-up reimagination of the system. The key innovation was the Evoformer, a novel neural network architecture that Jumper and his collaborators designed specifically for the protein folding problem.
The Evoformer processes two types of information simultaneously: a multiple sequence alignment (MSA) representation, which captures evolutionary relationships between related proteins, and a pair representation, which encodes the spatial relationships between every pair of amino acid residues. These two representations communicate with each other through alternating layers of attention mechanisms, allowing the network to reason jointly about evolutionary patterns and geometric constraints.
This architecture was deeply informed by Jumper's understanding of protein biophysics. The pair representation, for example, was designed to respect the triangular inequality — the geometric constraint that the distance between residue A and residue C cannot exceed the sum of the distances from A to B and B to C. By building this physical constraint directly into the network architecture, the Evoformer could produce geometrically consistent structures without the need for a separate optimization step.
The final structure was produced by a structure module that iteratively refined 3D coordinates, using an equivariant transformer that respected the rotational and translational symmetry of physical space. The entire system was end-to-end differentiable — meaning it could be trained directly to produce accurate structures, without the brittle two-stage pipeline of AlphaFold 1.
The CASP14 Triumph
When AlphaFold 2 entered CASP14 in November 2020, the results were extraordinary. The system achieved a median GDT score of 92.4, with many predictions essentially indistinguishable from experimentally determined structures. On some targets, AlphaFold 2's predictions were more accurate than low-resolution experimental structures.
For Jumper, the moment of realization came before the official results were announced. As the team analyzed their predictions against the hidden experimental structures, they could see that the system was producing structures of unprecedented accuracy. "There was a period where we were just checking result after result, and they were all correct," Jumper later recalled. "At some point you realize that something fundamental has changed."
The CASP organizers agreed. John Moult, the founder of CASP, declared that the protein structure prediction problem had been "largely solved." It was a statement that would have been unthinkable just two years earlier.
The Nobel Prize
In October 2024, John Jumper and Demis Hassabis were awarded the Nobel Prize in Chemistry, sharing the honor with David Baker of the University of Washington, who was recognized for his complementary work in computational protein design. Jumper, then 39, became one of the youngest Nobel laureates in Chemistry in recent decades.
The prize recognized not just AlphaFold's accuracy, but its impact. By solving protein structure prediction, Jumper and his colleagues had removed a bottleneck that had constrained biology for half a century. The implications extended across every branch of the life sciences, from understanding disease mechanisms to engineering new enzymes to designing better gene editing tools.
Open-Sourcing AlphaFold
One of the most consequential decisions the team made — with strong support from both Jumper and Hassabis — was to open-source AlphaFold 2's code and make its predictions freely available. In partnership with EMBL-EBI, DeepMind released the AlphaFold Protein Structure Database, which grew to contain predicted structures for over 200 million proteins, covering nearly every known protein sequence.
The decision democratized structural biology overnight. Researchers in developing countries who had no access to cryo-electron microscopy or X-ray crystallography facilities could now obtain high-quality protein structures for free. Within a year of its release, the database had been accessed by over a million researchers worldwide, and AlphaFold predictions had been cited in thousands of scientific publications.
Impact on Gene Editing
AlphaFold's impact on gene editing has been direct and substantial. CRISPR-Cas gene editors are proteins, and understanding their three-dimensional structures is essential for improving their performance. With AlphaFold, researchers can model the structures of novel Cas proteins discovered in metagenomic surveys, predict how mutations will affect their activity and specificity, and design engineered variants with improved properties.
The ability to rapidly predict protein structures has accelerated the development of next-generation gene editors — including smaller Cas proteins that are easier to deliver into cells, high-fidelity variants with reduced off-target effects, and entirely new classes of editors discovered through computational screening of natural protein diversity.
AlphaFold 3 and Future Directions
After AlphaFold 2, Jumper continued to lead the development of AlphaFold 3, released in 2024, which extended the system's capabilities beyond single-chain protein structure prediction to model the structures of complexes involving proteins, DNA, RNA, small molecules, and ions. This was a critical advance, because most biological processes involve interactions between multiple types of molecules.
AlphaFold 3's ability to model protein-DNA interactions is particularly relevant for gene editing, where the precise geometry of the Cas protein-guide RNA-target DNA complex determines editing efficiency and specificity.
Recent Developments (2025–2026)
Following his Nobel Prize in Chemistry 2024, Jumper has been recognized with the 2025 Golden Plate Award of the American Academy of Achievement, the Marshall Medal, and was elected Fellow of the Royal Society (FRS). In 2026, he was elected to the National Academy of Engineering.
Now serving as Director at Google DeepMind, Jumper's next research direction focuses on fusing AlphaFold's deep but narrow protein prediction capabilities with the broad reasoning of large language models. "I'll be shocked if we don't see more and more LLM impact on science," he has stated. His team is exploring how to integrate LLMs with structural biology tools for enhanced scientific discovery.
Research Lab & Companies
- Google DeepMind — Director (AlphaFold team lead)
- Nobel Prize in Chemistry 2024 — shared with Demis Hassabis (and David Baker)
- Fellow of the Royal Society (FRS) — elected 2025
- National Academy of Engineering — elected 2026
Jumper continues to work at DeepMind, pushing the boundaries of what AI can contribute to biology. His trajectory — from physics graduate student struggling with the limitations of molecular dynamics to Nobel laureate who solved one of biology's greatest challenges — is a testament to the power of bringing the right combination of deep domain knowledge and computational innovation to bear on the right problem at the right time.
