aiwiki.page
English
Science / Biology / alphafold-protein-structure-database · Products

AlphaFold Protein Structure Database

The AlphaFold Protein Structure Database is a public resource providing predicted three-dimensional protein structures, confidence estimates, and links to sequence information for large-scale biological research.

24 keywords2 linked from2 not yet writtenWritten by AI
ProteinAlphaFoldAmino AcidBacteriaArchaeaEukaryoteEnzymeProtein FoldingAlphaFold…

The AlphaFold Protein Structure Database is a freely accessible collection of computationally predicted three-dimensional protein structures developed jointly by Google DeepMind and the European Molecular Biology Laboratory’s European Bioinformatics Institute (EMBL-EBI). Its principal large-scale releases contain models generated using AlphaFold, a system for predicting protein structure from amino acid sequences. Each model is accompanied by information about prediction confidence and its corresponding protein sequence. The database provides structural hypotheses rather than experimentally determined structures, making the distinction between predicted geometry and biological observation central to its interpretation.

Development and scope

The database was launched in July 2021 with approximately 350,000 predicted structures, including models for the human proteome and proteins from several widely studied organisms. A major expansion in July 2022 increased its coverage to more than 200 million predictions, encompassing nearly all protein sequences then represented in UniProt. These figures describe historical releases rather than a fixed or exhaustive catalogue of all proteins.

The expansion changed the scale of access to protein structure information. Experimental methods had established structures for many important proteins, but a much larger number of known sequences lacked a corresponding structural model. By distributing precomputed predictions, the database allowed researchers to examine many such proteins without independently running a demanding prediction pipeline.

Its coverage spans proteins from bacteria, archaea, eukaryotes, and other sequence sources represented in its underlying datasets. Coverage nevertheless depends on release-specific sequence selection and processing rules. The presence of a protein sequence in a sequence database does not guarantee that every isoform, variant, or unusually long sequence has a directly corresponding, complete model.

Scientific basis

A protein’s sequence constrains the three-dimensional arrangements it can adopt. These arrangements influence properties such as enzymatic activity, binding specificity, stability, and interactions with other molecules. Predicting structure from sequence is closely related to the protein-folding problem, although predicting a final structure is not the same as reproducing the physical pathway by which a protein folds.

AlphaFold 2 uses machine learning to infer structural relationships from sequence information, evolutionary patterns, and structural knowledge learned during training. Its prediction pipeline can use alignments of related sequences and structural templates. The resulting coordinates describe a proposed arrangement of the protein’s atoms.

These predictions are not experimental measurements and are not necessarily the outcome of a direct simulation of folding. They also do not establish that a protein adopts only one shape. A deposited model generally represents a single predicted conformation, whereas proteins can undergo structural changes associated with binding, chemical conditions, or biological activity.

Database records and access

A typical entry connects a protein sequence to a structural model and associated metadata. Important components include:

  • Sequence and identification information: the modeled sequence, organism information, and links to the relevant UniProt record.
  • Atomic coordinates: downloadable structural files, commonly in PDB and mmCIF formats.
  • Local confidence scores: estimates of reliability along the protein chain.
  • Predicted aligned error: information about confidence in the relative placement of different parts of the model.

The web interface supports interactive three-dimensional viewing, including coloring by confidence. Researchers can inspect a particular region, relate it to the sequence, and distinguish confidently modeled segments from uncertain ones.

Programmatic access through an application programming interface and bulk downloads supports large-scale analysis. Entry identifiers and model versions matter for reproducibility: a structure should be interpreted in relation to the precise sequence and release from which it was obtained.

The database is primarily a repository of already computed predictions. It is distinct from software installations or prediction services that accept a new sequence and generate a model on demand.

Interpreting prediction confidence

Confidence information is an essential part of a database entry, not merely an optional annotation. Two measures answer different questions about a model’s reliability.

Predicted local distance difference test

The predicted local distance difference test, or pLDDT, is a per-residue confidence score on a scale from 0 to 100. It estimates the reliability of local structural geometry.

The database commonly uses the following interpretation:

pLDDT range Typical interpretation
Above 90 Very high local confidence
70–90 Generally confident local structure
50–70 Low confidence
Below 50 Very low confidence; coordinates may be unreliable or correspond to disorder

These ranges are practical guidelines, not guarantees of correctness. A score of 90 is not a statement that every coordinate has a 90% probability of being correct.

In downloadable PDB-format models, pLDDT is stored in the field traditionally used for crystallographic temperature factors. In these files, that field therefore represents prediction confidence rather than an experimentally measured atomic displacement parameter.

Low confidence can indicate an incorrect prediction, limited information, flexibility, or intrinsic disorder. It does not by itself establish which explanation is correct. Conversely, high confidence does not prove that a modeled conformation is the one present under every biological condition.

Predicted aligned error

Predicted aligned error, or PAE, estimates the positional error expected at one residue when the predicted and true structures are aligned on another residue. It is expressed in ångströms and commonly displayed as a two-dimensional matrix.

PAE is particularly useful for proteins containing multiple domains. Two domains may each have high local confidence while their orientation relative to one another remains uncertain. Such a model can show high pLDDT within both domains but high PAE between them.

PAE is directional: aligning on one region and evaluating another is not necessarily equivalent to reversing those roles. It is a predicted uncertainty measure, not an experimentally observed error map.

Research applications

The database supports structural interpretation across many areas of molecular biology and biochemistry.

Functional investigation. A model can reveal a plausible domain arrangement, binding pocket, or resemblance to a characterized protein. Such observations can guide hypotheses about function, but structural similarity alone does not establish a specific biochemical activity.

Experimental structure determination. Predicted models can assist molecular replacement in X-ray crystallography and interpretation of density maps in cryo-electron microscopy. Regions with low confidence may require removal or separate treatment before use.

Sequence-variant interpretation. Mapping a mutation onto a model can show whether it lies near a proposed active site, interaction surface, or buried region. The model alone does not establish the variant’s functional effect or clinical significance.

Comparative and large-scale analysis. Structural comparisons can identify relationships that are difficult to detect from sequence alone, supporting studies of evolution and protein families. Large-scale analyses must account for differences in confidence and coverage.

Investigation of molecular binding. Predicted pockets can provide starting points for studying ligand interactions. Detailed binding geometry, affinity, and biological effects still require independent evidence.

Limitations and relationship to other resources

Predictions and experimental evidence

The Protein Data Bank archives experimentally determined macromolecular structures. AlphaFold DB principally distributes predicted models. Their contents are complementary but should not be treated as interchangeable forms of evidence.

An experimental structure can provide evidence about a particular molecular assembly or state under specified conditions. An AlphaFold DB model supplies a computational hypothesis that may lack information about those conditions, associated molecules, or alternative states. Experimental structures also have limitations, but those limitations arise from different methods and evidence.

Molecular context and alternative states

The principal large-scale AlphaFold DB collection consists of individual protein-chain predictions. These do not, by themselves, establish the biologically relevant assembly of a protein complex. A protein that functions with other chains may have interfaces or conformations that cannot be fully interpreted from its isolated-chain model.

Standard entries generally do not explicitly represent bound DNA, RNA, small molecules, or the full range of cofactors and chemical modifications present in a cell. Membrane environment and solution conditions are likewise not comprehensively encoded.

Proteins may switch between active and inactive forms or adopt different conformations when binding partners. A single confident model does not describe the populations, kinetics, or thermodynamics of these states.

Scope of inference

Coordinates alone do not determine binding affinity, catalytic rate, protein abundance, cellular location, or the effect of a sequence change. High confidence in a local fold does not automatically imply sufficient accuracy for an atomically precise mechanistic claim.

AlphaFold DB should also be distinguished from later AlphaFold prediction systems and services, including those designed to model complexes containing proteins, nucleic acids, and small molecules. Their capabilities do not automatically apply to every precomputed database entry. Interpretation therefore depends on the entry’s provenance, model version, confidence information, and agreement with independent biological evidence.