AlphaFold creator John Jumper on turning a 50-year biology challenge into a 2-minute AI computation

Summary

John Jumper, the DeepMind researcher behind AlphaFold, makes a simple case for what AI can do that humans structurally cannot: compress a year of experimental biology into two minutes of computation.

The problem AlphaFold solved — predicting how a protein folds from its amino acid sequence — had resisted every computational approach for roughly 50 years. The experimental alternative was brutal. Scientists first had to produce enough pure protein to study, a process that could stall for months before any real work began. Then they had to coax the protein into forming an artificial crystal structure, with no reliable method for doing so. One paper Jumper cites notes that "after more than a year, crystals began to grow" — meaning researchers discovered something from a year-old cabinet experiment had quietly worked. After crystallization, the structure still had to be resolved at a synchrotron. The full process cost around $100,000 and a year of effort, with wide error bars on both.

AlphaFold 2 eliminated that process. The system runs in minutes and achieves near-experimental accuracy, and DeepMind made it openly available. Applications now range from vaccine design to filling gaps in the map of human biology.

What made AlphaFold 2 work

The breakthrough was not better data — everyone in the field used the same 140,000-structure protein data bank that scientists had collectively built over 50 years. It was data efficiency. Jumper says the team treated the problem as machine learning research rather than an application problem, rebuilding core model components inspired by the transformer but purpose-built for protein structure. An external group later confirmed the gap: AlphaFold 2 trained on just 1% of available data still outperformed the original AlphaFold system trained on all of it. That is a roughly 100-fold improvement in data efficiency between versions one and two.

The project's origins were scrappier than the outcome might suggest. Early work explored reinforcement learning and energy minimization — approaches borrowed from AlphaGo — but Jumper says those were wrong directions. A parallel thread started when a small internal group used a hackathon week to Google "grand challenges in biology" and found protein folding on the list. Both threads converged, and the team stayed with the problem through multiple blind alleys before the data-driven supervised learning approach finally worked.

Market reaction and commercial translation

Biotech public markets barely moved when AlphaFold was announced, and Jumper's explanation is structural rather than a sign of missed impact. A drug costs roughly $1 billion in R&D, so eliminating a $100,000 step is not the gap to commercial value. The more relevant question is whether better structural predictions raise clinical trial success rates — trials are the dominant cost driver, and they currently fail more than 90% of the time. If AI tools let researchers select the right experiment in the first or third attempt rather than the hundredth, the clinical cost curve moves.

The market signal that Jumper points to is in private valuations. Companies including Evolutionary Scale and Xaira have raised large rounds on the back of AI-enabled protein design, which sits downstream of AlphaFold. The public market didn't have a clean pure-play vehicle at launch in the way CRISPR companies provided in that wave — there was no direct equivalent of Editas or Intellia for protein structure prediction.

FDA and the role of computation

On regulatory adaptation, Jumper is careful to stay in his lane but makes one structural point: AI tools change where researchers apply their judgment, not the standards of evidence regulators require. Safety and efficacy still need real-world clinical evidence. What changes is that researchers have much better predictions going into experiments, which should mean fewer wasted trials before finding the right molecule or dose. That framing sidesteps the question of whether the FDA needs to accept computational outputs as evidence — it doesn't have to, if AI simply makes the experiments that do generate evidence more likely to succeed.

AlphaFold 3 extended the same approach to protein-small molecule binding, which is directly relevant to drug design. Alphabet's Isomorphic Labs is the commercial vehicle pursuing that direction.