The peculiarity of “overnight breakthroughs” is how silent they appear at the time. Not a parade. Don’t yell in public. A small group of people realizing the numbers in front of them aren’t a coincidence, someone’s half-drunk coffee cooling next to a keyboard, or the late-night glow from lab monitors… That was the atmosphere surrounding AlphaFold’s performance at CASP14, a biannual competition with the vigor of an academic sports league, except that the stakes are essentially modern biology and the scoreboard is protein structures.
The protein folding problem has been a sort of running dare for about half a century. Proteins are known to be chains of amino acids. With the help of physics, chemistry, and the messy reality of water, we are aware that those chains consistently collapse into complex three-dimensional shapes. And we are aware that shape determines fate: a functional enzyme or a biological catastrophe can result from the same sequence folded in a different way. The promise, which dates back to the 1970s, was straightforward to make but cruel to carry out: read the sequence and guess the shape.
| Item | Details |
|---|---|
| “Problem” | Predicting a protein’s 3D structure from its amino-acid sequence (the protein folding/structure prediction challenge) (Nature) |
| AI system | AlphaFold (Google DeepMind; later work with Isomorphic Labs on newer generations) (Google DeepMind) |
| Public benchmark moment | CASP14 (Critical Assessment of Structure Prediction), where AlphaFold achieved top performance (predictioncenter.org) |
| Reported performance detail | CASP14 median score around 92.4 GDT (often described as comparable to experimental accuracy for many targets) (Google DeepMind) |
| Why it matters | Protein shapes drive function; faster structure prediction speeds biology and drug research (Nature) |
| Where researchers look things up now | AlphaFold Protein Structure Database (AlphaFold DB) (alphafold.ebi.ac.uk) |
| One authentic reference | AlphaFold Protein Structure Database (EMBL-EBI + Google DeepMind) (alphafold.ebi.ac.uk) |
The old route was physically demanding and slow in practice. NMR, cryo-EM, and X-ray crystallography. They are all strong, picky, and reliant on costly equipment and patient people to manipulate proteins into acting in certain ways. Stories from structural biologists typically include long nights in beamline facilities, months of failed crystals, and a certain type of resigned humor. As you watch them work, you get the impression that science advances not just through genius but also through perseverance. This is why it felt like a line snapping taut when AlphaFold made accurate landing predictions that CASP organizers compared to lab methods.
The median score of 92.4 GDT, which is clean enough to fit on a slide, is part of DeepMind’s own explanation of CASP14. However, that number has a less neat human meaning. The system’s predicted structures frequently landed near experimental ones across targets, with the type of errors measured in angstroms—distances that seem fictitious until you realize they’re the scale of atoms. Perhaps the most disturbing aspect wasn’t that the AI “won,” but rather that it did so by a narrow margin; rather, it won by altering the definition of “good.”
If the word “trick” is correct, the fundamental trick wasn’t brute force. The combinatorial explosion has always been the nightmare of protein folding: a protein can theoretically take on an absurd number of different shapes before stabilizing. According to the peer-reviewed paper that followed, AlphaFold’s method relied on deep learning that had been trained on known structures to predict constraints and geometry in ways that allowed it to leap toward believable solutions rather than aim aimlessly. The fact that the model is generating incredibly helpful structures without replicating every microscopic stage of folding is the aspect that still causes some scientists to squint. Yes, it is useful. A complete description of how nature works? It’s still unclear.
Then came the second wave, which turned the breakthrough into infrastructure rather than a trophy. By 2022, a release of more than 200 million predicted structures connected to UniProt at the planetary scale was being described by DeepMind and EMBL-EBI. The experience of biology changed abruptly: you could look up a structure by scrolling and rotating it in a browser, just like people used to do with maps, rather than waiting for one. Nothing appeared different outside of Cambridge and other research centers; within labs, workflows subtly curved around this new default.
The story doesn’t end neatly, of course. Even AlphaFold fans will acknowledge that proteins can be erratic—changing shape, forming complexes, interacting with DNA, RNA, ions, medications, and membranes. Predictions are not experiments. The drive for younger generations has included closing that gap. AlphaFold 3 and a server designed to provide researchers with wider access for non-commercial use, emphasizing interactions among biomolecules rather than just single-protein shapes, were introduced in 2024 by Google DeepMind and Isomorphic Labs. The path is obvious. The line separating what is “solved” is constantly shifting.
It’s difficult to ignore the industry undertow that exists here. Investors seem to think that the headline is about what you can build on top of structural prediction, such as new wet-lab priorities, faster target validation, and drug discovery engines. As a reminder that scientific advances can swiftly become competitive moats, even Nature reported rumors in early 2026 about “AlphaFold 4“-like systems and drug-focused models that businesses might choose to keep confidential. There is an underlying tension between closed systems that feel like an arms race and open tools that feel like a gift.
It’s difficult not to experience two emotions at once while watching this play out. Relief, as a long-standing biological conundrum has been solved sufficiently to alter day-to-day tasks. And discomfort, as a small number of well-funded organizations are increasingly training, fine-tuning, and occasionally excluding the most promising scientific instruments. The need for experiments was not eliminated by AlphaFold. It made some experiments seem worthwhile, while others now seem, just a little, like walking when a train has begun to run alongside you.
