AlphaFold · AIxSci

What it is

Proteins are long chains of small units called amino acids. A cell makes the chain in a straight line, but the chain then folds into a compact three-dimensional shape, and that shape decides what the protein does. Working the shape out in the laboratory is slow and sometimes impossible. AlphaFold instead learns from the tens of thousands of protein shapes already solved by experiment, and uses what it has learned to guess the shape of a chain nobody has solved. It takes the sequence of amino acids as its input and returns a set of atom positions, together with a score saying how confident it is about each part.

The system does not fold anything in the way a cell does. It looks for patterns. One important input is a collection of related sequences from other species, gathered by searching large sequence databases: when two positions in a protein change together across species, they are often touching in the folded shape. AlphaFold combines these hints with any known structures of similar proteins and reasons about the whole chain at once, refining its answer in repeated passes. A companion version, called Multimer, predicts how two or more separate chains sit against each other. ColabFold is a lighter way of running the same models.

Why scientists use it

Scientists use it because shape is the shortest route to function. If you can see the folded form, you can guess which parts sit buried in the core and which are exposed, which pieces are likely to bind a drug, and which loops dangle loose. Before prediction, each of those questions needed months of crystallography or microscopy. A prediction arrives in hours, so it can be used as a starting guess, a sanity check on experimental data, or a way of choosing which proteins are worth the laboratory effort at all.

In the record

structural-biologyNeural network predicts protein atom positions from amino acid sequence aloneBuilt the network's inputs by searching sequence databases for related proteins and searching known structures for templates, alongside the protein's own sequence.structural-biologyResearchers measure folding stability for hundreds of thousands of protein variantsPredicted structures for candidate sequences with default settings, keeping the most confident model, so loosely attached ends could be trimmed before stability assays were designed.structural-biologyBenchmark measures how well AlphaFold predicts antibody-antigen complex structuresPredicted antibody and antigen bound together straight from sequence, using two released Multimer versions and repeated runs through ColabFold with different random starting points.

Limits

A prediction is a plausible shape, not a measurement. The system is most reliable for single, well studied, rigid proteins with many known relatives, and least reliable for flexible regions, rare sequences, and chains that only take shape when bound to something else. Its confidence score flags uncertainty but does not catch every error. It generally returns one fixed form, so it says little about how a protein moves, how it changes when a small molecule or a mutation arrives, or how strongly two proteins actually stick together.

Written by AIxSci: the background from general knowledge, the examples from checked records.