Reading AlphaFold DB: colors and confidence in predicted structures

Reading confidence in AlphaFold DB predictions: Nadoo Bio public database series

AlphaFold DB is a public database of three-dimensional protein structures predicted by the AI model AlphaFold. Its records describe how a protein may fold, calculated from the sequence of its small building blocks, amino acids. They offer a view of a protein's possible shape even when an experimental structure is not yet available.

Open a prediction and you will see blue and yellow regions. These colors carry information: they distinguish regions whose shape the model predicts with greater confidence from regions it finds harder to resolve. This article focuses on confidence in structures of individual proteins.

A prediction model and a structure database serve different purposes

AlphaFold is the name of a family of models that calculate structures. The AlphaFold Protein Structure Database, or AlphaFold DB, collects precomputed results. Opening a database record means looking up an existing prediction, rather than running a new calculation. The database provides both individual protein structures and complexes containing multiple molecules. [1]

A coordinate file contains predicted atom positions. Confidence information helps identify which regions can inform interpretation and which need additional evidence.

Connecting a sequence, a predicted structure, and a region of interest

Mapping a region of interest from the sequence onto the prediction lets you examine its shape and confidence together.

pLDDT: the confidence measure behind the colors

pLDDT measures local structural confidence for each residue, the amino-acid unit in a protein chain. Values range from 0 to 100. Higher values mean that the model has greater confidence in the local shape around that residue. AlphaFold DB displays these values as colors. [2]

pLDDT range

Confidence

Meaning when reading the structure

Above 90

Very high

The local shape is predicted with high confidence

70–90, inclusive

High

Consider the shape alongside other evidence

50 to below 70

Low

Interpret detailed atomic arrangements cautiously

Below 50

Very low

Consider whether the region is flexible or needs more information

Scores vary within the same structure. Flexible regions, or regions that do not readily adopt a single shape, may receive low scores. A low score alone does not establish the reason, but it signals that detailed atomic positions are difficult to interpret. [2]

pLDDT answers how confident the model is in a region's shape. A score of 90 does not mean a 90% experimental success rate. Binding strength, biological function, and drug effects require experimental evidence appropriate to each question.

Well-predicted parts can still have uncertain relative positions

A domain is a part of a protein that folds relatively independently. Imagine two domains joined by a thin linker. The model may predict each domain's shape well while remaining uncertain about the angle between them.

PAE, predicted aligned error, helps assess this relative placement. It is the model's estimate of how far one residue's position would differ after aligning the predicted and actual structures around another residue. It uses ångströms (Å), a unit for very small distances between atoms. Lower values indicate less uncertainty in relative position. This is a model estimate, not an error measured against an actual structure. [3]

pLDDT describes local shape confidence; PAE helps assess relative placement between parts

The shapes of individual parts can be clear while their relative placement remains uncertain. Reading both measures makes this distinction easier to see.

An analogy is checking the shape of individual toy blocks separately from the angles at which they connect. Proteins, however, move with their surroundings, so research must consider states beyond a single fixed arrangement.

Why the region of interest matters more than an overall average

Start with the region relevant to your question, rather than the average confidence of the entire structure. If you are interested in a site where a small molecule might fit, inspect nearby pLDDT values. To examine how two domains meet, read their pLDDT values together with PAE.

Match the sequence and the region of interest as well. Records with the same protein name can cover different sequences or regions, making it possible to compare different parts inadvertently. Experimental structures and functional studies can help interpret the predicted shape.

To investigate a bound state, you need data that includes the binding partner. The possible shape of one protein and the arrangement of multiple interacting molecules address different questions.

What a predicted structure can tell you

AlphaFold DB is useful because it presents a possible shape alongside confidence in that prediction. Reading local shape through pLDDT and relative placement through PAE helps define how far an interpretation can reasonably go.

A prediction read this way becomes a hypothesis for further experiments. As experimental evidence accumulates, it becomes clearer which parts agree with the prediction and which require other states to be considered.

Terms of use

AlphaFold DB prediction data may be used for both research and commercial purposes. The applicable CC BY 4.0 license permits copying, modification, and redistribution with attribution. Give the source and license link, and indicate any changes. [1] [4]

These terms refer to prediction data distributed by the database. AlphaFold code, model weights, and separate servers have their own terms. [1]

References

  1. AlphaFold Protein Structure Database. Frequently asked questions. Official help. Checked October 23, 2026.

  2. EMBL-EBI. pLDDT: Understanding local confidence. AlphaFold training material. Checked October 23, 2026.

  3. EMBL-EBI. PAE: A measure of global confidence in AlphaFold2 predictions. AlphaFold training material. Checked October 23, 2026.

  4. Creative Commons. Attribution 4.0 International (CC BY 4.0). License deed. Checked October 23, 2026.