AlphaFold 3 · AIxSci

What it is

Proteins are long chains built from a short alphabet of chemical units called amino acids. The chain folds up into a particular three-dimensional shape, and that shape decides what the protein does. Working the shape out in the laboratory is slow and sometimes impossible. AlphaFold 3 instead learns the link between sequence and shape from the thousands of structures biologists have already measured, then applies what it has learned to a new sequence. Given the sequence, it returns coordinates for every atom, together with a score saying how confident it is about each part.

The earlier generations handled proteins on their own. The third generation widens the scope: it can predict a protein together with the other things it grips in a living cell, such as DNA, RNA, or a small molecule that binds in a pocket. It also predicts several chains arranged around each other, so a pair of proteins can be modelled as the pair rather than one at a time.

Why scientists use it

Most questions in biology and medicine turn on shape. Whether a mutation matters, whether a drug will fit a pocket, how two proteins lock together: all of these are easier to reason about with a model of the atoms in front of you. Prediction stands in for an experiment that has not been done, and often for one that cannot easily be done at all. Researchers use the predicted structures as a starting point, comparing them with each other, with measured structures, or with versions of the same protein from different species.

Because the predictions are quick and need only a sequence, they also serve as raw material for further computation. A predicted structure can be fed into another program, fitted into laboratory measurements, or used to pick out which parts of a molecule to study more closely. In each case the model is a hypothesis about shape, not a measurement of it.

In the record

structural-biologySoftware models of two cancer proteins used to rank mutations by likely effectBuilt full-length models of two human proteins, and of the same protein in yeasts, where no measured structures exist, then overlaid them to locate matching mutated positions.structural-biologyDeep learning reads cryo-EM maps alongside AlphaFold3 models to build protein structuresPredicted a structure for each protein chain from its sequence; the chains were then split into parts and fitted into electron microscope density maps.structural-biologyFive structure-prediction models tested on 509 protein–peptide pairs under altered inputsRun as one of several prediction programs on a test set of protein-peptide pairs, its outputs scored against the measured structures.

Limits

A prediction is a guess informed by known structures, not a measurement. Where a sequence resembles little that has been solved before, the guess can be confidently wrong, and the confidence scores do not catch every failure. Proteins also move, and a single set of coordinates says nothing about the shapes a molecule shifts between, or which one it takes inside a cell. Predicted contacts between a protein and another molecule are especially uncertain. The usual practice is to treat a model as a hypothesis and check it against experiment.

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