AlphaFold Confidence: Check How the Protein Regions Fit Together

A beautifully folded protein model can look finished. Two compact regions sit neatly beside each other, the colours suggest confidence, and a presentation caption announces that AI has revealed how the whole structure fits together. Before accepting that account, ask a more specific question: confidence in which part of the picture?

AlphaFold2 offers a useful lesson for anyone explaining AI science. Confidence in a region’s internal structure and confidence in its position relative to another region answer different questions. For an independent creator preparing an explainer, or a business owner reviewing a research pitch, that distinction can change the sentence the evidence supports.

Start with the model and the date

John Jumper and colleagues’ original AlphaFold paper was published online on 15 July 2021. This article examines AlphaFold2 confidence measures using EMBL-EBI’s training guidance. It is an explanation of established methods, rather than an announcement of a new 2026 result or a description of every later AlphaFold system.

That framing belongs near the beginning of a science story. A reader should know whether the author is discussing a particular model’s output, a newer release or an experimental finding. Keeping those identities visible gives subsequent claims somewhere precise to attach.

Local confidence has a local meaning

According to EMBL-EBI’s pLDDT guidance, predicted Local Distance Difference Test scores describe confidence at individual amino-acid positions, or residues. Scores run from 0 to 100. Higher values generally indicate greater confidence in the local structure; values above 90 fall in the very high confidence category.

A low score also needs interpretation. The guidance identifies different possibilities: a region may be flexible or intrinsically disordered, or the model may lack enough information to predict an otherwise structured region confidently. A low score alone does not establish which explanation applies.

For an editor, the useful question is where the confidence occurs. Ask the researcher to identify the particular region under discussion. A caption about one well-predicted section should name that section, rather than quietly extending its assurance to every feature in the image.

Relative placement needs a separate check

EMBL-EBI’s explanation of Predicted Aligned Error, or PAE, addresses relative positioning. PAE estimates the positional error at one residue if predicted and actual structures were aligned on another. It is expressed in ångströms, a unit of length.

Low PAE between regions indicates confidence in their relative placement; high PAE indicates uncertainty about it. The training guidance explains that pLDDT and PAE should be considered together. Strong local predictions can coexist with uncertain positioning of different domains, the structural regions within a protein.

This changes what an image can establish. If a proposed explanation depends on two regions being arranged together in a particular way, the reviewer should ask for evidence addressing that arrangement. The visual appeal of the rendered structure cannot supply the missing answer.

A fictional explainer shows the editorial decision

Imagine a creator preparing an educational animation about an invented protein with two domains. A researcher supplies a fictional AlphaFold2 output in which both domains have predominantly high local confidence, while their relative placement has high predicted aligned error. No actual protein, model run or experimental result is being reported here.

The animation brief proposes a close-up of the two domains meeting like a precisely fitted clasp. Its narration says that the model has established the clasp’s working arrangement. The problem is the added claim: the planned scene treats the uncertain relationship as settled.

The creator can revise the brief without abandoning the story. The animation could introduce each domain separately, then explain that the supplied prediction leaves their arrangement uncertain. If alternative arrangements are drawn for teaching purposes, they should be labelled as illustrations, with no suggestion that they are measured states or ranked model outputs.

The next discussion with the researcher becomes more useful: which arrangement is the project trying to understand, and what evidence would address it? The expert may have additional structural or experimental evidence. If so, that evidence deserves its own explanation, rather than being attributed to the confidence colouring.

This is an editorial proposal, not a validated research protocol. Its value is in making the change of claim visible before the animation turns an unsupported detail into its central visual.

Write the caption before polishing the graphic

A practical review can begin with three short statements: what the displayed object represents, which feature the accompanying evidence supports, and what remains unresolved. Drafting these first gives the illustrator a brief that includes the evidence boundaries.

In the fictional project, a suitable caption would say that the two domains have confident local predictions while their relative arrangement remains uncertain in the supplied output. It would avoid implying that the imagined clasp had been observed working. This wording also gives the audience a reason to care about the next investigation.

Keep a small source record alongside the production brief:

• The model or database entry, its version or creation date, and the exact output used.
• The researcher’s description of the regions and confidence measures relevant to the claim.
• The evidence supporting any proposed relationship between the regions.
• The caption and narration approved for the final illustration.

These are proposed production records. They help a later revision stay connected to the material actually reviewed. If a replacement output arrives, revisit the explanation as well as swapping the image.

Match the explanation to the question

A related distinction appears in MaryChuks.com’s guide to what an AlphaGenome DNA prediction represents: a molecular output needs a clear account of the claim being made from it. Here, the question is specifically about structural confidence at different scales.

For the fictional creator, the final decision is concrete. Show the confident regional predictions, explain the unresolved arrangement, and ask the researcher what would justify a stronger account. A compelling science story can take the reader to the edge of the evidence and make the remaining question understandable.

Sources

Primary sources checked on 29 October 2026. The example is fictional. The EMBL-EBI training pages do not display publication dates; the paper’s date refers to its first online publication.

• Jumper and colleagues: Highly accurate protein structure prediction with AlphaFold — Nature, 15 July 2021.
• EMBL-EBI: pLDDT — Understanding local confidence.
• EMBL-EBI: PAE — A measure of global confidence in AlphaFold2 predictions.