pLDDT and PAE: interpreting AlphaFold structure confidence

Two metrics answer different questions

For AlphaFold2 models, pLDDT gives per-residue local confidence on a 0–100 scale. It helps assess how reliably a region is described. A high value concerns expected geometry, not protein thermostability.

PAE, predicted aligned error, is expressed in angstroms and describes expected positional error when aligning on another model region. Individual domains may be confidently predicted while their relative positions remain uncertain.

What Foldiff displays

The current beta retrieves existing AlphaFold DB structures by accession. It does not run AlphaFold. In 3D, choose cartoon, surface or sticks, overlay a homolog and highlight candidates, differences, deviations, confidence or functional sites.

The view retains the original reference structure. Highlighting a candidate helps inspect its environment, but does not show a newly calculated mutant structure. A geometric overlay does not establish equal activity either.

Why interpret PAE together with pLDDT?

EMBL-EBI recommends interpreting these metrics together. When domain orientation confidence is low, interpret inter-domain contacts cautiously. The PAE diagonal alone says little: each residue is aligned with itself.

Foldiff’s PAE mode helps explore structure blocks. It does not fit to experimental cryo-EM density or establish domain boundaries. Interpretation also needs the sequence, functional annotations and research question.

A short checklist before choosing substitutions

1. Match the accession, length and original amino acid to your construct.
2. Inspect confidence at the substitution position and its surroundings.
3. Check whether the conclusion depends on uncertain relative domain positions.
4. Inspect functional annotations; missing annotations do not guarantee safe intervention.
5. Save the structure, report and calculation parameters for later review.

These steps reveal limits before ordering variants. If a suitable entry is missing, the current beta should not imply that it has reliably calculated a structure on demand.

Start with a worked example

On the examples page you can explore subtilisin and lysozyme. Switch highlights, select a position and compare the reference with a homolog. This demonstrates public-entry analysis, not a published client pilot result.

Confidence metrics help select hypotheses. Comparable laboratory measurements establish stability and retained function.