AI in Science: AlphaFold
AlphaFold is an artificial intelligence system developed by DeepMind that predicts the three-dimensional shape of proteins from their amino acid sequences — a problem that had stumped biologists for over fifty years. It matters because protein shape determines protein function, and understanding that link is foundational to medicine, drug discovery, and our understanding of life itself. The remarkable fact: AlphaFold solved in months what the entire global scientific community had been unable to crack in half a century, and then DeepMind gave its predictions away for free to every researcher on Earth.
Every protein is a long chain of chemical building blocks called amino acids, and that chain folds itself into a precise three-dimensional shape — the same shape, every single time — determined entirely by the sequence of those building blocks. The shape is everything: it decides what the protein can do, what it can grab onto, and whether a drug molecule will fit it like a key in a lock. AlphaFold's insight was that this folding problem, which chemists had attacked with physics and equations, could instead be treated as a pattern-recognition problem — and that modern deep learning, trained on the library of already-known protein structures, could learn the hidden rules of folding that nature had been using for billions of years.
How It Works
To understand how AlphaFold works, it helps to think about how you might recognize a face you have never seen before. You do not compute every bone angle or skin texture from scratch — you draw on millions of past faces you have seen, and your brain finds the pattern that fits. AlphaFold does something similar, but for molecular shapes.
Step 1 — Reading the sequence. AlphaFold starts with a string of letters representing the amino acid sequence of a protein. Each letter stands for one of twenty possible chemical building blocks, like beads on a necklace.
Step 2 — Searching evolutionary history. The system searches enormous databases of protein sequences collected from thousands of different species. Because evolution preserves what works, two amino acids that always appear together across millions of years of evolution are almost certainly physically close to each other in the folded protein. AlphaFold mines this co-evolution signal as a treasure map of spatial relationships.
Step 3 — Building a relationship map. Using a powerful neural network architecture called a transformer — the same family of models behind modern language AI — AlphaFold builds a rich table of predictions about which amino acids are near which other amino acids in 3D space, even before it knows the full shape.
Step 4 — Iterative refinement. The system passes its predictions back through itself multiple times, each pass refining its understanding, like a sculptor who keeps stepping back to look at the whole figure before adding another detail.
Step 5 — Outputting a structure. Finally, AlphaFold produces a complete 3D coordinate map of every atom in the protein, along with a confidence score for each region — the system genuinely knows which parts of its answer it is sure about and which parts are uncertain. The key innovation was combining evolutionary signals, transformer-style attention across the amino acid sequence, and iterative self-correction all in one end-to-end trained system.
John Jumper trained as a computational chemist and physicist before joining DeepMind, bringing rare expertise in both the biology of protein folding and the engineering of large neural networks. He led the team that built AlphaFold 2 — the version that stunned the world at the CASP14 competition in 2020 — designing the novel Evoformer architecture that processes evolutionary and structural information simultaneously. In 2024, Jumper shared the Nobel Prize in Chemistry for this work, becoming one of the youngest Nobel laureates in the prize's history and the first to win it primarily for developing an AI system.
Demis Hassabis is a former child chess prodigy, video game designer, and neuroscientist who founded DeepMind in 2010 with the explicit goal of using artificial intelligence to accelerate scientific discovery. He championed the protein folding problem as a grand challenge for AI, arguing years before AlphaFold's success that deep learning could unlock it. In 2024, Hassabis shared the Nobel Prize in Chemistry alongside Jumper — an extraordinary outcome for someone trained primarily as a computer scientist and cognitive neuroscientist rather than a chemist.
Intuition
Imagine you have never seen a folded origami crane, but you have a book of a thousand photographs showing completed cranes alongside the precise sequence of folds that made each one. After studying enough examples, you begin to notice patterns: certain sequences of early folds always lead to the same later shapes; certain crease combinations never appear together in a finished crane. Given a new fold sequence you have never seen, you could make a surprisingly good guess about what the finished shape looks like — not by physically folding paper, but by pattern-matching to everything you have learned. AlphaFold does exactly this, but the "fold instructions" are amino acid letters, and the "shapes" are the three-dimensional sculptures those letters produce. There is one additional trick that makes it more powerful: evolution has been running experiments for three billion years. If the same stretch of amino acids appears in a fish, a mushroom, and a bacterium all doing the same job, evolution has essentially confirmed that that sequence folds into a stable, functional shape. AlphaFold reads this ancient experimental record and uses it as a guide. The confidence scores AlphaFold reports are like a good student flagging which parts of their exam they are sure of and which parts they guessed on. Regions of a protein that are floppy or variable in real life show up as low-confidence predictions — the system is not just right, it also knows when it might be wrong, which is a hallmark of genuinely useful scientific tools.
Historical Development
Origins · 1 minhow this was discovered
Why It Matters
Prerequisites
What This Leads To
See Also
Transformer Architecture · Deep Learning · Reinforcement Learning from Human Feedback · AI for Scientific Discovery · Generative Models in Biology