Highly accurate protein structure prediction with AlphaFold

Inside the paper

The problem

Predicting a protein’s three-dimensional structure from its sequence remained difficult, particularly when no closely related structure had already been solved. Physical simulation and methods based on evolutionary history each faced substantial limits.

The idea

AlphaFold combines information from related sequences and structural templates in a network that predicts atomic coordinates. Its representations exchange information repeatedly before the final structure is constructed.

How it works

• A database search supplies a multiple sequence alignment and, where available, candidate structural templates for the target sequence.
• The Evoformer updates representations of aligned sequences and pairs of residues, allowing evolutionary and geometric information to inform each other.
• A structure module turns these representations into a spatial arrangement of residues and predicts atomic positions.
• The model passes its own outputs back through the network, refining the prediction through recycling, and estimates local confidence.

What it showed

In the blind CASP14 assessment, the paper reports a median backbone error of 0.96 Å on assessed domains, compared with 2.8 Å for the next best method. Accuracy also varied with the depth of the available sequence alignment.

Read it for

Start with Figure 1 for the overall flow and the comparison with experimental structures, then read the architecture discussion around Figure 4. Figure 5 is useful for understanding where sequence information helps; the detailed methods can wait until you want to implement the model.

Written in our own words from the paper itself. A draft the editors have not yet reviewed.