AlphaFold Protein Structure Database
Google DeepMind and EMBL-EBI
More than 260 million predicted protein structures, roughly a thousand times what experiments have solved.
AlphaFold DB holds AI-predicted 3D structures for nearly every protein in UniProt, each with per-residue confidence scores (pLDDT) and predicted aligned error. Curated collections highlight pathogens with pandemic potential, antimicrobial resistance proteins, and WHO priority proteomes. Biologists use it as a starting structure where no experimental one exists.
Access
- Cost
- FreeFree web, API, FTP, and Google Cloud bulk copies.
- License
- Open licenseCC BY 4.0
- Sign-up
- No sign-up
- How you get it
- API and Bulk downloadmmCIF, PDB, JSON
What's in it
- One row is
- One predicted structure per UniProt protein
- Covers
- Not geographic; organisms across the tree of life
- Years
- 2021-present
- Updated
- Irregular
- Size
- Over 260 million predictions; full download is tens of terabytes
Quality
High accuracy for well-folded domains, with honest per-residue confidence. Predictions are single chains and miss ligands, complexes, and conformational changes.
- Low-confidence regions are often disordered, not wrong
- No complexes or bound ligands
Connects to
- Protein Data Bank (PDB): Experimental structures used for training and validation
- UniProtKB (UniProt Knowledgebase): Predictions keyed on UniProt accessions
- gnomAD (Genome Aggregation Database): Place human missense variants on predicted structures
Questions it could answer
- What share of each species' proteome is predicted with high confidence?
- Do disease-causing missense variants cluster in confidently folded regions?