AlphaFold structure prediction
/ AL-fuh-fold /
For half a century biologists faced a maddening gap: a protein's amino-acid sequence is easy to read, but its three-dimensional shape — the thing that actually determines what it does — was painfully slow and expensive to determine, taking months or years of laboratory work per protein. The dream was to predict the shape directly from the sequence, by computer. AlphaFold, a program from the company DeepMind, is the breakthrough that finally brought that dream within reach for a huge fraction of proteins.
AlphaFold is an artificial-intelligence system trained on the tens of thousands of protein structures that had been painstakingly solved by experiment. From a query sequence it does two clever things: it gathers many related sequences from other organisms (because positions that change together across evolution are often physically close in the folded shape), and it uses a neural network to weigh how every amino acid relates to every other, then iteratively refines a predicted set of atomic positions. The output is a full three-dimensional model, and crucially it comes with a per-residue confidence score, so a user can see which parts of the prediction to trust and which to treat with caution. In a 2020 community contest it produced models close enough to experiment that the field declared the long-standing prediction problem largely solved for many cases.
The impact has been enormous: AlphaFold and its open database have provided predicted structures for nearly every known protein, accelerating drug discovery, enzyme engineering, and basic research, and its creators shared a 2024 Nobel Prize. But honesty matters here. Solved is an overstatement if taken literally. AlphaFold predicts a single most-likely folded shape; it is far weaker at how proteins move, how they change shape when binding partners, disordered regions that have no fixed fold, the effect of single mutations, and how proteins assemble with others. A confident-looking model is a hypothesis, not a measurement — it still needs experimental checking for anything important.
A researcher studying a poorly known bacterial protein pastes its sequence into the AlphaFold database and gets back a 3D model in seconds, color-coded by confidence. The well-folded core is bright with high confidence; a floppy tail is dim, warning that this region has no reliable predicted shape.
Seconds of computing replace months of crystallography — with confidence flags attached.
AlphaFold predicts one likely static shape, not motion, not the effect of a single mutation, and not disordered regions well. A high-confidence model is a strong hypothesis, not a proven structure.