AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models — TEI-tagged view

Mihaly VaradipersName, Stephen AnyangopersName, Mandar DeshpandepersName, Sreenath NairpersName, Cindy NatassiapersName, Galabina YordanovapersName, David YuanpersName, Oana StroepersName, Gemma WoodpersName, Agata LaydonpersName, Augustin ŽídekpersName, Tim GreenpersName, Kathryn TunyasuvunakoolpersName, Stig PetersenpersName, John JumperpersName, Ellen ClancypersName, Richard GreenpersName, Ankur VorapersName, Mira LutfipersName, Michael FigurnovpersName, Andrew CowiepersName, Nicole HobbspersName, Pushmeet KohlipersName, Gerard KleywegtpersName, Ewan BirneypersName, Demis HassabispersName, Sameer VelankarpersName (2021date). “AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models.” Nucleic Acids Researchtitle·j.

Object of study — the AlphaFold Protein Structure Databaseterm, its predicted structures, confidence outputs, and access mechanisms

Research question — how to expand protein-sequence-space structural coverageterm and deliver high-accuracy predicted structures with interpretable confidence information

Contribution — an openly accessible database initially providing more than 360,000 AlphaFoldterm structures, confidence estimates, and bulk, APIterm, and interactive access

Full text — tagged

AlphaFold Protein Structure Databaseterm: massively expanding the structural coverageterm of protein-sequence spaceterm with high-accuracy models Abstract The AlphaFold Protein Structure Databaseterm (AlphaFold DBterm, https://alphafoldterm.ebi.ac.uk ) is an openly accessible, extensive databaseterm of high-accuracy protein-structure predictionsterm. Powered by AlphaFoldterm v2.0 of DeepMind, it has enabled an unprecedented expansion of the structural coverageterm of the known protein-sequence spaceterm. AlphaFold DBterm provides programmatic access to and interactive visualization of predicted atomic coordinatesterm, per-residue and pairwise model-confidence estimatesterm and predicted aligned errorsterm. The initial release of AlphaFold DBterm contains over 360,000 predicted structures across 21 model-organism proteomes, which will soon be expanded to cover most of the (over 100 million) representative sequences from the UniRef90term data set.

INTRODUCTION Proteins are essential macromolecules with vital biological functions and, thus, are involved in a wide range of research activities and medical and biotechnological applications, from fighting infectious diseases to tackling environmental pollution ( 1 , 2 ). Knowledge of the three-dimensional (3D) arrangement of the atoms of a protein can provide essential clues to understanding the roles and mechanisms underpinning protein functions ( 3 , 4 ). However, while the Universal Protein Resourceterm (UniProtterm) archives almost 220 million unique protein sequences, the Protein Data Bankterm (PDBterm) holds only just over 180 000 3D structures for over 55 000 distinct proteins, thus severely limiting the coverage of the sequence space to support biomolecular research globally ( 5–7 ).

Achieving a higher coverage of the sequence space with experimentally determined high-resolution structures is very labour-intensive. It often requires a lot of trial and error, for example, to find suitable constructs or conditions under which a protein is amenable to crystallization. Although recent advances in the field of electron cryo-microscopy and hybrid and integrative methods (I/HM) for structure determination have accelerated the pace of structure determination, the gap between known protein sequences and experimental protein structures continues to expand ( 6 , 8 ).

One way to close this gap is to predict the structures of millions of proteins. Increasingly, researchers deploy Artificial Intelligenceterm (AI) techniques to predict a protein's structure computationally from its amino-acid sequenceterm·var alone ( 9–11 ).

AlphaFoldterm is an AI systemterm developed by DeepMind that makes state-of-the-art predictions of protein structures from their amino-acid sequencesterm·var ( 9 ). CASPterm (Critical Assessment of Structure Predictionsterm) is a biennial challengeterm for research groups to test the accuracy of their predictions against actual experimental data. In 2020, the organizers of the CASP14 benchmark recognized AlphaFoldterm as a solution to the protein–structure–prediction problem ( 12 ). The unprecedented accuracy and speed of AlphaFoldterm allowed the creation of an extensive database of structure predictions at a large scale. It will enable biologists to obtain structural models for almost any protein sequence, changing how they tackle research questions and accelerate their projects. The methodology of AlphaFoldterm and insights gained from the predictions for the complete human proteome have been described recently ( 9 , 13 ).

We present the AlphaFold Protein Structure Databaseterm (AlphaFold DBterm, https://alphafoldterm.ebi.ac.uk ), a new data resource created in partnership between DeepMind and the EMBL-European Bioinformatics Institute (EMBL-EBI). We have created AlphaFold DBterm to make structure predictions freely available to the scientific community at a large scale. The first release described here covers the human proteome and those of 20 other model organisms (Table 1 ). In the coming months, we plan to have expanded the database to cover a large proportion of all catalogued proteins (over 130 million cluster representatives from UniRef90term).

Structural predictions for complete proteomes in AlphaFold DBterm IMPLEMENTATION The initial version of AlphaFold DBterm contains over 360 000 predicted structures, corresponding meta-information and confidence metrics. All the data are publicly accessible through a cloud-based infrastructure. We have attempted to predict most sequences in the UniProt reference proteometerm in the 16–2700 amino acid length range (as well as 1400-residue fragments to cover longer human proteins) for the organisms currently covered. We excluded sequences that contain non-standard amino acids. We do not provide multiple isoforms at this point.

The predicted structures contain atomic coordinates and per-residue confidence estimatesterm on a scale from 0 to 100, with higher scores corresponding to higher confidence. This confidence measureterm is called pLDDTterm and corresponds to the model's predicted per-residue scoresterm on the lDDT-Cα metric ( 14 ). lDDTterm is a pre-existing metric used in the protein structure prediction field. A key motivation behind lDDTterm is to assess the local accuracy of a predictionterm, awarding a high score for regions that are well-predicted even if the entire prediction cannot be aligned well to the true structure. This is particularly important for evaluating multi-domain predictions where the individual domains may be largely accurate while their relative position is not. As a confidence metric based on lDDTterm, pLDDTterm also reflects local confidence in the structure, and should be used, for example, to assess confidence within a single domain. Several other protein structure prediction resources also use lDDT-based metrics ( 15 , 16 ). AlphaFold DBterm stores these values in the B-factor fields of the mmCIF and PDBterm files available for download and uses confidence bands based on these values to colour-code the residues of the models in the 3D structure viewer on the structure pages. Residues with pLDDTterm ≥ 90 have very high model confidenceterm, while residues with 90 > pLDDTterm ≥ 70 are classified as confident. Residues with 70 > pLDDTterm ≥ 50 have low confidence, and residues with pLDDTterm < 50 correspond to very low confidence ( 13 ). It was recently described that very low confidence pLDDT scoresterm correlate with high propensities for intrinsic disorderterm ( 17 ).

The Predicted Aligned Errorterm (PAEterm) is another output of the AlphaFoldterm system. It indicates the expected positional errorterm at residue x if the predicted and actual structures are aligned on residue y (using the Cα, N and C atoms). PAEsterm are measured in Ångströms and capped at 31.75 Å. Scientists can use these values to assess the confidence in the relative position and orientation of different parts of the model (e.g. two domains). For residues x and y in two different domains, if the PAE valuesterm·var ( x , y ) are low, AlphaFoldterm predicts the domains to have well-defined relative positions and orientations. If the PAE valuesterm·var are high, then the relative position and orientation of the two domains are unreliable, and users should not attach biological or structural relevance to these. Note that the PAEterm is asymmetric; therefore, there can be a difference between the PAE valuesterm·var for ( x, y ) and ( y, x ), for example, between loop regions with highly uncertain orientation.

Data archival AlphaFold DBterm archives and provides access to the atomic coordinates in PDBterm and mmCIF formats, PAEsterm in JSON format and corresponding metadata in JSON format. While the coordinates and the PAEterm files are directly accessible through URLs, we load and index the metadata using the open-source search platform Apache Solrterm ( https://solr.apache.org/ ) to enable users to search on the AlphaFold DB web pagesterm. The data files in the archive are versioned, and previous snapshots of the data will be available via FTP, but the web pages will always display the latest version.

Data access AlphaFold DBterm provides predictions through several data-access mechanisms: (i) bulk downloads via FTP; (ii) programmatic access via an application programming interfaceterm (APIterm); (iii) download and interactive visualization of individual predictions on protein-specific web pages keyed on UniProt accessionsterm.

For bulk downloading data from AlphaFold DBterm, users can access the uncompressed archive files (.tar) of compressed PDBterm/mmCIF files (.gz) per reference proteome from the EMBL-EBI public FTP area at ftp://ftp.ebi.ac.uk/pub/databases/alphafoldterm . This area contains the TAR files and a JSON file that provides meta-information, describing the species names (scientific and common), the reference proteome identifiers, the number of predicted structures, and the archives’ sizes. The same information and files are also available from the Bulk Download page of AlphaFold DBterm at https://alphafoldterm.ebi.ac.uk/download .

We provide access to all entries through a public APIterm endpoint, keyed on a UniProt accessionterm. For example, the endpoint https://alphafoldterm.ebi.ac.uk/apiterm/prediction/Q92793 allows access to all the meta-information and the URLs of all the archived data files related to the human CREB-binding protein. UniProtterm ( 5 ), Pfam ( 18 ), InterPro ( 19 ) and PDBe-KB ( 7 ) use this APIterm to display AlphaFoldterm models on their web pages.

AlphaFold DBterm provides graphical access to and interactive visualization of all the predictions and meta-information for the broader scientific community through web pages. These pages contain all the available information for a protein of interest, keyed by its UniProt accessionterm. They allow users to analyse the prediction and download the corresponding model files (in PDBterm and mmCIF formats) and PAEterm files (in JSON format).

AlphaFold DB web pagesterm AlphaFold DBterm provides convenient access to its predictions through a set of web pages ( https://alphafoldterm.ebi.ac.uk ). These pages contain an introduction to the AlphaFoldterm system, address the most frequent questions, enable bulk download of complete proteomes, and offer a search engine for finding pages specific to a protein of interest (Figure 1 ). Users can search by gene name, protein name, UniProt accessionterm or organism name. The search results can be filtered, for example, only to show human proteins.

Searching AlphaFold DBterm. AlphaFold DBterm provides a search engine to find proteins of interest based on gene or protein name, UniProt accessionterm or organism name. The search results can be filtered if required and clicking on a protein name leads to the relevant protein-specific entry page.

Each protein has a dedicated structure page that shows basic information (drawn from UniProtterm ( 5 ) and PDBe ( 6 )) and three separate outputs of the AlphaFoldterm model. The first two outputs are the 3D coordinates and the per-residue confidence metric pLDDTterm, which is used to colour the residues of the model in the integrated 3D molecular viewerterm, Mol* ( 20 ). Model confidenceterm can vary significantly along a chain, making it essential to analyse the confidence measuresterm before interpreting structural features. The lower confidence bands appear to correlate well with backbone flexibility and intrinsic disorder ( 13 ) (Figure 2 ).

Meta-information and 3D visualization of the AlphaFoldterm structure predictions. The protein-specific web pages display essential metadata for the protein of interest, such as known biological functions and cross-references to UniProtterm and PDBe-KB. Users can download the predicted models in PDBterm and mmCIF format, and an interactive molecular viewer visualizes the structure, coloured by the per-residue pLDDTterm confidence measureterm.

The third output is a pairwise confidence predictionterm, which helps to assess the reliability of relative domain positions and orientationsterm as well as the global topology of the protein (Figure 3 ). The plot is coloured by the pairwise PAE valuesterm·var and it helps users to identify which domains have reliably predicted positions and orientations relative to one another, where dark greenpersName indicates high confidence. Selecting a region in the plot also highlights the corresponding part of the sequence in the 3D viewer.

Visualization of Predicted Aligned Errorsterm. Protein-specific pages contain an interactive 2D plot of the PAE valuesterm·var. This tool interacts with the 3D molecular viewerterm to facilitate the identification of domains whose relative positions and orientations AlphaFoldterm predicts with confidence. In this example ( https://alphafoldterm.ebi.ac.uk/entry/Q93074 ), AlphaFoldterm has high confidence in the relative position of domains at residues 1–500 (greenpersName) and residues 1200–1700 (blue), but not with the region between 500–1200 (orange) nor the C-terminus.

CONCLUSION AND OUTLOOK Since the mid-1950s, the scientific community has been using ever-more advanced experimental methods to determine over 180 000 structures of proteins, nucleic acids, and complexes in atomic detail, and archive them in the PDBterm, the single worldwide archive of experimental macromolecular structure data managed by the wwPDB consortium ( 21 ). This collective body of work has vastly improved our understanding of many fundamental processes in health and disease, as evidenced in part by many Nobel Prizes for structures deposited in the PDBterm. Recently, determining the structure of the SARS-CoV-2 viral proteins enabled scientists to understand how it operates and to identify potential treatments and develop new vaccines ( 3 ). However, figuring out the exact structure of a protein remains an expensive and often time-consuming process. Thus, we only know the 3D structure of a tiny fraction of all proteins currently known to science.

The first release of AlphaFold DBterm contains over 360 000 predicted structures from 21 model-organism proteomes. Having access to these highly accurate models will greatly impact biology, from enabling structure-based drug design to providing data for high-throughput structural bioinformatics researchterm that will tackle fundamental biological questions. We have already gained some invaluable insights from the predictions of the human proteome ( 13 ).

In the coming months, we will expand AlphaFold DBterm to provide structural predictions to include additional proteomes to support research in neglected diseases and to cover the set of highly annotated proteins in SwissProt, taking the number of structures available to >1 million. This will be followed by another update in 2022 to include structures for most representative sequences from the UniRef90term data set (>100 million structures). Future updates will also aim to overlay annotations onto the predicted structures and display this information on 2D sequence-feature viewers. AlphaFold DBterm will enable biomedical scientists to use 3D models of protein structures as a core tool, driving research and innovation across multiple fields by providing open access to an ever-growing number of predicted structures.

DATA AVAILABILITY All the AlphaFoldterm predictions are publicly available through multiple data-access mechanisms. Coordinate files in PDBterm and mmCIF formats are available in TAR archives per proteome through FTP at ftp://ftp.ebi.ac.uk/pub/databases/alphafoldterm . Meta-information and URLs to individual UniProt accessionsterm are available via a public APIterm endpoint. For example, https://alphafoldterm.ebi.ac.uk/apiterm/prediction/Q92793 provides all the information for UniProt accessionterm Q92793 ( https://www.alphafoldterm.ebi.ac.uk/entry/Q92793 ).

ACKNOWLEDGEMENTS We would like to acknowledge all the scientists who contributed valuable feedback throughout the development of this data resource. We would also like to recognize the contributions of all the structural biologists whose experimentally determined structuresterm, archived in the PDBterm, enabled the training of AlphaFoldterm. We further acknowledge the work of the public protein sequence archives such as the UniProtterm consortium, BFD and MGnify in collecting and organizing protein-sequence data which was used for predicting structures.

Contributor Information Mihaly VaradipersName, European Molecular Biology Laboratory, European Bioinformatics Institute, Hinxton, UK.

Stephen AnyangopersName, European Molecular Biology Laboratory, European Bioinformatics Institute, Hinxton, UK.

Mandar DeshpandepersName, European Molecular Biology Laboratory, European Bioinformatics Institute, Hinxton, UK.

Sreenath NairpersName, European Molecular Biology Laboratory, European Bioinformatics Institute, Hinxton, UK.

Cindy NatassiapersName, European Molecular Biology Laboratory, European Bioinformatics Institute, Hinxton, UK.

Galabina YordanovapersName, European Molecular Biology Laboratory, European Bioinformatics Institute, Hinxton, UK.

David YuanpersName, European Molecular Biology Laboratory, European Bioinformatics Institute, Hinxton, UK.

Oana StroepersName, European Molecular Biology Laboratory, European Bioinformatics Institute, Hinxton, UK.

Gemma WoodpersName, European Molecular Biology Laboratory, European Bioinformatics Institute, Hinxton, UK.

Agata LaydonpersName, DeepMind, London, UK.

Augustin ŽídekpersName, DeepMind, London, UK.

Tim GreenpersName, DeepMind, London, UK.

Kathryn TunyasuvunakoolpersName, DeepMind, London, UK.

Stig PetersenpersName, DeepMind, London, UK.

John JumperpersName, DeepMind, London, UK.

Ellen ClancypersName, DeepMind, London, UK.

Richard GreenpersName, DeepMind, London, UK.

Ankur VorapersName, DeepMind, London, UK.

Mira LutfipersName, DeepMind, London, UK.

Michael FigurnovpersName, DeepMind, London, UK.

Andrew CowiepersName, DeepMind, London, UK.

Nicole HobbspersName, DeepMind, London, UK.

Pushmeet KohlipersName, DeepMind, London, UK.

Gerard KleywegtpersName, European Molecular Biology Laboratory, European Bioinformatics Institute, Hinxton, UK.

Ewan BirneypersName, European Molecular Biology Laboratory, European Bioinformatics Institute, Hinxton, UK.

Demis HassabispersName, DeepMind, London, UK.

Sameer VelankarpersName, European Molecular Biology Laboratory, European Bioinformatics Institute, Hinxton, UK.

FUNDING Funding for open access charge: DeepMind.

Conflict of interest statement . None declared.

REFERENCES Associated Data Data Availability Statement All the AlphaFoldterm predictions are publicly available through multiple data-access mechanisms. Coordinate files in PDBterm and mmCIF formats are available in TAR archives per proteome through FTP at ftp://ftp.ebi.ac.uk/pub/databases/alphafoldterm . Meta-information and URLs to individual UniProt accessionsterm are available via a public APIterm endpoint. For example, https://alphafoldterm.ebi.ac.uk/apiterm/prediction/Q92793 provides all the information for UniProt accessionterm Q92793 ( https://www.alphafoldterm.ebi.ac.uk/entry/Q92793 ).