TrustCurve

AlphaFold gives every residue a confidence score, pLDDT, from 0 to 100. Is a 90 worth more than a 60? Check it against experiments: take structures measured after the predictions were public, lay each one on its prediction, and count, band by band, how many residues landed close.

the Sketcher's drawing Prediction Computed by a model, with how sure it is. Not an experiment.

the Photographer's picture Experiment Measured from real molecules.

The data: 2991 residues from 12 proteins, the BlindTest issues. A band with fewer than 30 residues says "too few to say".

Your guess first

Of the residues AlphaFold scored very high (90 and above), how many sat within 2 Å of the experiment after the two were laid on each other? (2 Å is a little more than the length of one chemical bond.)

How it is counted

Each protein's experiment and AlphaFold DB model are paired residue by residue (UniProt numbering) and laid on each other by a least-squares fit (the Kabsch method) on the residues the model scored 70 or more.
Each residue's distance is between its CA atoms (one per residue). Its band is the model's own pLDDT: very high 90 and above, confident 70 to 90, low 50 to 70, very low below 50.
The proteins: Alpha-globin transcription factor CP2 (PDB 8Y7V), L-methionine gamma-lyase (PDB 9GJ9), D-aspartate oxidase (PDB 9VPF), Proton-activated chloride channel (PDB 9HQN), Antiviral protein MAP (PDB 9X2O), Transmembrane protein 164 (PDB 9LW1), Type 3 secretion system ATPase (PDB 9DMD), L-erythrulose-1-phosphate isomerase (PDB 11EF), Nudix hydrolase 11 (PDB 9NXM), T-cell antigen CD7 (PDB 12HP), Chromosomal replication initiator protein DnaA (PDB 10XE), C5a anaphylatoxin chemotactic receptor 2 (PDB 9V35).

Honest notes

This is a handful of proteins, each a single chain, and a handful is not a survey.
The fit uses the confident residues, which favours them a little. Crystal contacts, partners, membranes and flexible loops can move a residue for reasons that are not the model's fault.
pLDDT scores local shape. Whether two parts sit at the right angle is what PAE describes, so a confident part can still be "far" when the whole chain is fitted at once.
The real reference is the large-scale checks in the AlphaFold papers, done on many more structures with other measures (lDDT): Jumper et al. 2021 and Tunyasuvunakool et al. 2021.

Sources

Jumper et al. 2021, Highly accurate protein structure prediction with AlphaFold, Nature 596:583 (what pLDDT is, and how it was checked).
Tunyasuvunakool et al. 2021, Highly accurate protein structure prediction for the human proteome, Nature 596:590.
Experimental structures: RCSB Protein Data Bank. Predictions: AlphaFold Protein Structure Database (CC BY 4.0).
The superposition: Kabsch 1976, Acta Cryst. A32:922.

Snapshot 2026-10-08.