HonestReport

1. "Predicted" means computed, not measured

Most structures in the AlphaFold Database are computer predictions made from a protein's sequence. An experimental structure (from X-ray crystallography, cryo-electron microscopy or NMR) is measured from real molecules. A prediction can be very accurate, but it is still a model, and it comes with its own estimate of where it might be wrong.

2. pLDDT: how sure the model is about each residue

Every residue gets a score from 0 to 100. Above 90 is very high (the backbone and most side chains are usually placed well); 70 to 90 is confident (the backbone is usually right); 50 to 70 is low; below 50 is very low and should not be read as a real shape on its own.

A low score does not mean "unimportant" or "broken". Very low stretches are often parts of the protein that are flexible or disordered in real life, and many of them do real jobs.

3. PAE: how sure the model is about where parts sit relative to each other

The predicted aligned error is a map over pairs of residues, in ångströms. Low error between two parts means the model is sure how they sit relative to each other. A model can be very sure about each folded domain and still unsure how the domains are arranged, so a claim about two distant parts needs the PAE, not just the pLDDT colours.

4. What a single-chain model leaves out

A standard AlphaFold Database entry is one protein chain on its own. It does not show partner proteins, DNA, drugs, metal ions or other small molecules bound to it, and it shows one shape where the real protein may move between several. A prediction of the normal sequence also does not show, on its own, what a mutation does.