When people talk about what AI has given science, one name comes up most often. To see why, we have to start with proteins.
Why a protein's shape matters
Almost everything that happens in a living cell is done by proteins. Haemoglobin carries oxygen, enzymes break down food, antibodies catch viruses, insulin tells cells to take sugar from the blood. Every protein is a long chain of amino acids, twenty kinds of building block. The order of the blocks is called the sequence, and we have been able to read it for decades.
But a protein does not work as a stretched-out chain. Picture a beaded necklace that folds up in space into an intricate tangle. The order of the beads is the sequence; the resulting shape of the tangle is the structure. And it is the shape that decides what the protein does: which molecule it fits, what it can cut, where a drug can bind. Working out the shape experimentally, for instance by X-ray crystallography, took months to years of work per protein. So for half a century biologists asked: could the shape be computed straight from the sequence?
What AlphaFold does
AI gave the answer. AlphaFold is a family of models from DeepMind (now Google DeepMind) that predicts the three-dimensional structure from the sequence and from information about related proteins. In 2020 the second generation, AlphaFold 2, reached an accuracy at the CASP14 assessment that, for many proteins, approached experimental measurements. CASP is a regular blind test: teams receive the sequences of proteins whose shapes labs have already measured but not yet published, and the predictions are compared to see who comes closest. Details are in the Nature paper.
Terms that often get mixed up:
- AlphaFold 2 (2020) is the model that made protein-structure prediction famous.
- AlphaFold 3 (2024) goes further: it also predicts the joint structures of complexes where a protein meets DNA, RNA or a small molecule such as a drug. It is not just a renamed AlphaFold 2. AlphaFold 3 paper.
- AlphaFold DB is the public database of finished predictions that Google DeepMind maintains with EMBL-EBI, the European bioinformatics institute. It holds over 200 million structures and labs around the world use it. AlphaFold DB.
What the Nobel Prize recognised
In 2024 the Nobel Prize in Chemistry recognised two different approaches to proteins. Half went to David Baker for the computational design of new proteins that do not exist in nature. The other half was shared by Demis Hassabis and John Jumper of Google DeepMind for protein-structure prediction — that is, for AlphaFold. The same year, the Nobel Prize in Physics went to John Hopfield and Geoffrey Hinton for the foundations of machine learning with neural networks; 2024 showed that AI had become a tool of science, not just its subject. Official announcement.
How to read a prediction
A colourful picture of a molecule looks convincing, but it is an estimate with varying confidence. AlphaFold attaches a confidence score to every part of the protein: pLDDT says how sure the model is about the local shape of a segment, PAE estimates how accurately that segment is placed relative to another part. A low-confidence segment is not necessarily the model's mistake; some parts of proteins simply have no fixed shape. Neither number says anything about how well a drug will act on the protein. Guide to quality measures.
It also matters that AlphaFold predicts the final shape, not the route to it. It does not say how the chain folds step by step inside the cell, nor how the protein moves and changes shape while working. So the word fold in the name does not mean the protein-folding puzzle is solved. Strengths and weaknesses are summarised in the EMBL-EBI training guide.
From structure to drug is a long way
An example: a pharmaceutical team is looking for a substance that blocks a protein a virus needs in order to multiply. AlphaFold shows the protein's shape and a pocket where a small molecule might bind. That saves months and suggests which substances to try. Whether the substance really binds, whether it stops the virus, whether it harms patients and at what dose — all of that still has to be shown by the lab and by clinical trials. AlphaFold solves one well-defined, important step, not the whole of drug development.
And that is also a lesson for the superintelligence debate. AlphaFold is narrow AI in the best sense: in one field it surpassed decades of human work, and it can do nothing else. It shows how large AI's contribution to science can be, and also that a top result in one field does not amount to general intelligence. Where AlphaFold sits among other systems is discussed on AI, AGI and ASI.