ReferenceAI skillsStructure prediction and foldingReferences

Structural confidence versus measurement

Check input requirements and the meaning of structural confidence when predicting structures from protein sequences or reviewing folding results.

Skill version: 1.1.0. Resource: skill://nadoobio-folding/references/confidence.md.

pLDDT is the model's prediction confidence in local structure at each residue. Higher generally means greater confidence, but check the output contract for its range. pLDDT 90 does not mean a 90% binding success rate.

PAE is the expected positional error in one part after alignment on another, usually in Å with lower being better. It is not measured coordinate error. ipTM is confidence in relative chain arrangements and does not measure affinity or function.

Comparing only overall averages across candidates with different sequences, chains, or lengths can hide errors in important regions. Compare equivalent regions and conditions and explain flexible regions separately. Experimental validation remains necessary even with high confidence.

From input to reporting

Materials to checkHow to read themWhen materials are missing
Input sequences/chains and output coordinatesCheck sequence correspondence and missing positions for each chain. Full sequence length may differ from the number of residues with coordinates.Do not guess chain names or missing coordinates.
Per-residue confidence, averages and scaleRead the average separately from values in the region of interest, using the range declared by that output.Do not report an accurate binding site based on an average alone.
Repeated candidates and selection criteriaDistinguish variation across all supplied candidates from the selected candidate’s values.Do not infer uniform quality across the run from the highest score alone.

Hypothetical teaching example: on a 0–100 scale, an overall average of 90 with a value of 50 in the region of interest supports the report, “Overall mean confidence is high, but the region of interest needs further assessment.” Here, 50 is an example, not a universal exclusion threshold. Next, check per-residue values, chain correspondence and any needed reference structure. A request for this explanation does not itself authorize rerunning.

Source: Structure prediction methods, confidence and comparison with experimental structures. The numbers above illustrate interpretation and are not paper results. Read the current server contract to determine support in a particular public tool.

On this page