AlphaFold Protein Structure Database: Access, Download and Confidence Scores

The AlphaFold Protein Structure Database (AlphaFold DB), run by EMBL-EBI, holds predicted three-dimensional models for proteins, each keyed to a UniProt accession. This guide covers how to find a model, how to download it from the web page or the API, and how to read its confidence scores before you use it. Details were checked against the live service and EMBL-EBI training material on 2026-10-29.

What it is used for in research

Researchers use AlphaFold DB to get a structural model of a protein for which no experimental structure is at hand: to plan constructs, interpret mutations, prepare molecular replacement or docking inputs, or simply to see how domains may be arranged. Every model is a prediction, not a measurement. EMBL-EBI’s training course states that each structure comes with confidence metrics that support critical assessment of the prediction, and those metrics are what separate a usable region from a speculative one.

Experimentally determined structures live elsewhere. For those, see our guides to the Protein Data Bank and to X-ray crystallography. Sequence and functional annotation for the same protein is in UniProt, and UniProt accessions are the key AlphaFold DB uses.

Install or access (official source only)

There is nothing to install. The database is open in a browser at alphafold.ebi.ac.uk. According to EMBL-EBI’s training page, you can search by protein name, UniProt accession number or protein sequence, and the search box also takes gene and organism names. An entry page shows the 3D view, the per-residue confidence (pLDDT), the predicted aligned error (PAE) plot, a link to UniProt, function information and a structural cluster. The web pages always display the latest model version; earlier versions are available through the FTP archive.

Step-by-step: find and download a model

Through the web page

1. Open alphafold.ebi.ac.uk and search by protein name, gene, organism, UniProt accession or sequence.
2. Choose the entry whose accession matches the protein you mean. For a human protein, confirm the organism before you go on.
3. Look at the 3D view coloured by confidence and at the PAE plot before downloading anything.
4. Download the coordinates. EMBL-EBI lists PDB and mmCIF as the coordinate formats; PAE data is provided as JSON.

Through the API

If you already hold UniProt accessions, request the prediction record and read the file addresses from it, rather than building file names by hand.

1. Request https://alphafold.ebi.ac.uk/api/prediction/<UniProt accession>. The response is a JSON list with one record per model.
2. Read cifUrl, bcifUrl or pdbUrl for the coordinates, paeDocUrl for the PAE matrix, plddtDocUrl for per-residue pLDDT, and amAnnotationsUrl for the amino-acid substitution annotations file.
3. Fetch the file from the address you were given.

Worked example, queried on 2026-10-29 for human haemoglobin subunit alpha, accession P69905. The record returned the entry ID AF-P69905-F1, latestVersion 6, a model created date of 2025-08-01, and these addresses:

• Coordinates: https://alphafold.ebi.ac.uk/files/AF-P69905-F1-model_v6.cif (also .bcif and .pdb).
• PAE: https://alphafold.ebi.ac.uk/files/AF-P69905-F1-predicted_aligned_error_v6.json.
• pLDDT: https://alphafold.ebi.ac.uk/files/AF-P69905-F1-confidence_v6.json.

The version number is part of the file name, and it was 6 on the date checked. Older tutorials show _v4 in the same pattern; those file names are out of date. Reading the address from the API record means a script keeps working when the version changes. The same record reported a global confidence value of 98.06 for this model, with 99.3 percent of residues in the very high pLDDT band.

Output and how to read it

pLDDT: how confident is each residue?

pLDDT is a per-residue measure of local confidence on a 0 to 100 scale, where higher means more confident and usually more accurate. EMBL-EBI’s course describes a score above 90 as the highest accuracy category, with backbone and side chains typically predicted well, and a score above 70 as usually a correct backbone with some misplaced side chains. Regions below 50 can reflect natural flexibility or intrinsic disorder, or a region where the model lacks information to predict a structure confidently. The score can vary a lot along one chain, so judge the structure region by region instead of by the global value.

PAE: how well are two parts placed relative to each other?

PAE (predicted aligned error) shows confidence in the relative position of two residues. The plot is a square grid with a dark diagonal. pLDDT does not tell you whether two domains are positioned correctly with respect to one another, and PAE does, so read the two together.

A reading workflow

1. Check the organism and UniProt accession against the protein you intend to study.
2. Colour by pLDDT and mark the confident cores and the low-confidence tails or loops.
3. Open the PAE plot. If two confident domains show high error between them, treat the orientation between them as unreliable even though each domain looks good.
4. Record the entry ID, model version and the date you downloaded the file, so your data management plan and methods section can identify exactly which model you used.
5. Describe the model as predicted in any figure, table or deposition. Do not present it as an experimental structure.

Troubleshooting

• The API returns an error for your identifier. The path takes a UniProt accession. A malformed identifier returns a 400 response with the message “Invalid identifier format”, so check the accession against UniProt before suspecting the service.
• A download link from an older paper or script fails. File names carry the model version. Request the prediction record again and use the addresses it returns.
• Parts of the model are poorly defined. Look at the pLDDT. Very low scores often mark flexible or disordered regions, so a loose tail is not a software fault.
• Domains look misplaced relative to each other. Open the PAE plot. High error between domains means their arrangement is uncertain.
• A web page and a downloaded file seem to disagree. The web pages show the latest version. A file you saved earlier may be a previous version.

Free alternatives and companions

Resource	What it holds	Use it when
Protein Data Bank	Experimentally determined structures	You need a measured structure rather than a prediction
UniProt	Sequences and functional annotation, and the accessions AlphaFold DB uses as keys	You need to confirm which protein and isoform you are looking at

FAQ

Is an AlphaFold model the same as an experimental structure?

No. Each AlphaFold DB entry is a prediction, and the confidence metrics (pLDDT and PAE) exist so that you can critically assess it. For experimental structures use the Protein Data Bank.

How do I search the AlphaFold database?

By protein name, UniProt accession or protein sequence, and also by gene or organism name, on alphafold.ebi.ac.uk.

Which file formats can I download?

Coordinates are available as mmCIF and PDB, and the API record also lists a binary CIF (.bcif) address. PAE data is JSON.

Why does my file name say v6 when a tutorial says v4?

The model version is part of the file name, and the version returned by the API on 2026-10-29 was 6. Use the address in the API record.

Sources (official documentation)

Checked 2026-10-29.

• AlphaFold Protein Structure Database, EMBL-EBI: https://alphafold.ebi.ac.uk/. Prediction API record for P69905 (https://alphafold.ebi.ac.uk/api/prediction/P69905), checked 2026-10-29.
• EMBL-EBI Training, Accessing predicted protein structures in the AlphaFold Database: https://www.ebi.ac.uk/training/online/courses/alphafold/accessing-and-predicting-protein-structures-with-alphafold/accessing-predicted-protein-structures-in-the-alphafold-database/, checked 2026-10-29.
• EMBL-EBI Training, pLDDT: Understanding local confidence: https://www.ebi.ac.uk/training/online/courses/alphafold/inputs-and-outputs/evaluating-alphafolds-predicted-structures-using-confidence-scores/plddt-understanding-local-confidence/, checked 2026-10-29.
• EMBL-EBI Training, PAE: a measure of global confidence in AlphaFold predictions: https://www.ebi.ac.uk/training/online/courses/alphafold/inputs-and-outputs/evaluating-alphafolds-predicted-structures-using-confidence-scores/pae-a-measure-of-global-confidence-in-alphafold-predictions/, checked 2026-10-29.