Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 77 of the 83 references it cites were found in a registry or an index; 18 of the 24 sources its claims cite were checked for support.

Highly accurate protein structure prediction for the human proteome

Open access

Abstract

Protein structures can provide invaluable information, both for reasoning about biological processes and for enabling interventions such as structure-based drug development or targeted mutagenesis. After decades of effort, 17% of the total residues in human protein sequences are covered by an experimentally determined structure 1 . Here we markedly expand the structural coverage of the proteome by applying the state-of-the-art machine learning method, AlphaFold 2 , at a scale that covers almost the entire human proteome (98.5% of human proteins). The resulting dataset covers 58% of residues with a confident prediction, of which a subset (36% of all residues) have very high confidence. We introduce several metrics developed by building on the AlphaFold model and use them to interpret the dataset, identifying strong multi-domain predictions as well as regions that are likely to be disordered. Finally, we provide some case studies to illustrate how high-quality predictions could be used to generate biological hypotheses. We are making our predictions freely available to the community and anticipate that routine large-scale and high-accuracy structure prediction will become an important tool that will allow new questions to be addressed from a structural perspective.

Overview

CoChat Summary

TL;DR

AlphaFold2 was applied to 98.5% of human proteins, producing confident predictions for 58% of residues and very high confidence for 36% of all residues. The paper introduces pLDDT, pTM and an experimentally resolved head to interpret the dataset, presents case studies, and releases predictions freely.

Key findings

  • The dataset covers 98.5% of human proteins with full-chain predictions, amounting to 10,537,122 residues (92.5% of residues); 58.0% of residues were predicted confidently (pLDDT > 70), and 35.7% of total residues fell in the highest accuracy band (pLDDT > 90), double the residues covered by experimental structures.
  • At the per-protein level, 43.8% of proteins had a confident prediction on at least three quarters of their sequence, and 1,290 proteins contained a substantial region (>200 residues) with pLDDT ≥ 70 and no good template.
  • On a held-out set of long PDB chains (n = 151), 70% of predictions had a TM-score > 0.7; pTM correlated with actual TM-score at Pearson’s r = 0.84, and among human proteins enriched for novel multidomain predictions, 187 had pTM > 0.8 and 343 had pTM > 0.7.
  • pLDDT performed as a competitive disorder predictor on the CAID benchmark with AUC 0.897, while the experimentally resolved head achieved AUC 0.921; case studies proposed mechanistic hypotheses for G6Pase-α, DGAT2 and wolframin.

Methods

  • AlphaFold2 was applied to the UniProt human reference proteome (release 2021_02), attempting all sequences of 16–2,700 amino acids and excluding residue codes B, J, O, U, Z or X; the pipeline used MSA construction, template search, inference with five models, model ranking by mean pLDDT, and constrained relaxation, with Reduced BFD and reduced ensembling.
  • Evaluation used a recent PDB dataset of 12,494 chains, CAMEO targets from 21 March 2020 to 13 March 2021, and the CAID Disprot-PDB benchmark; metrics included pLDDT, pTM, the experimentally resolved head, TM-score, lDDT-Cα and AUC.
  • Case-study analyses used PyMOL, P2Rank, AutoDockTools, AutoDock Vina and TM-align; inference for the human proteome used 930 GPU days, with an estimated 510 core days for other pipeline stages from scratch.

Limitations

  • The paper states that low-confidence regions may reflect intrinsic disorder or regions structured only in complex, and that some proportion of proteins without confident predictions will be genuine failures where a fixed structure exists but AlphaFold does not predict it.
  • The paper states that inter-domain accuracy is expected to be lower than within-domain accuracy, that some multidomain proteins may have mobile domains with no fixed packing, and that the case-study biological hypotheses require experimental confirmation.
  • Evident from the text: predictions are single-chain and exclude sequences longer than 2,700 residues, so complex-dependent structure and larger proteins are not directly addressed.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41586-021-03828-1

Cites

77 of 77 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Model confidence and added coverage. a, Correlation between per-residue pLDDT and lDDT-Cα. Data are based on a held-out set of recent PDB chains (Methods) filtered to those with a reported resolution of <3.5 Å (n = 10,215 chains and 2,756,569 residues). The scatterplot shows a subsample (1% of residues), with the blue line showing a least-squares linear fit and the shaded region a 95% confidence interval estimated with 1,000 bootstrap samples. The black line shows x = y, for comparison. The smaller plot is a magnified region of the larger one. On the full dataset, the Pearson’s r = 0.73 and the least-squares linear fit is y = (0.967 ± 0.001) × x + (1.9 ± 0.1). b, AlphaFold prediction and experimental structure for a CASP14 target (PDB: 6YJ1)64. The prediction is coloured by model confidence band, and the N terminus is an expression tag included in CASP but unresolved in the PDB structure. c, AlphaFold model confidence on all residues for which a prediction was produced (n = 10,537,122 residues). Residues covered by a template at the specified identity level are shown in a lighter colour and a heavy dashed line separates these from residues without a template. d, Added residue-level coverage of the proteome for high-level GO terms, on top of residues covered by a template with sequence identity of more than 50%. Based on the same human proteome dataset as in c (n = 10,537,122 residues).
Fig. 1 Model confidence and added coverage. a, Correlation between per-residue pLDDT and lDDT-Cα. Data are based on a held-out set of recent PDB chains (Methods) filtered to those with a reported resolution of <3.5 Å (n = 10,215 chains and 2,756,569 residues). The scatterplot shows a subsample (1% of residues), with the blue line showing a least-squares linear fit and the shaded region a 95% confidence interval estimated with 1,000 bootstrap samples. The black line shows x = y, for comparison. The smaller plot is a magnified region of the larger one. On the full dataset, the Pearson’s r = 0.73 and the least-squares linear fit is y = (0.967 ± 0.001) × x + (1.9 ± 0.1). b, AlphaFold prediction and experimental structure for a CASP14 target (PDB: 6YJ1)64. The prediction is coloured by model confidence band, and the N terminus is an expression tag included in CASP but unresolved in the PDB structure. c, AlphaFold model confidence on all residues for which a prediction was produced (n = 10,537,122 residues). Residues covered by a template at the specified identity level are shown in a lighter colour and a heavy dashed line separates these from residues without a template. d, Added residue-level coverage of the proteome for high-level GO terms, on top of residues covered by a template with sequence identity of more than 50%. Based on the same human proteome dataset as in c (n = 10,537,122 residues).CC BY · PubMed Central
Fig. 2 Full chain structure prediction. a, TM-score distribution for AlphaFold evaluated on a held-out set of template-filtered, long PDB chains (n = 151 chains). Includes recent PDB proteins with more than 800 resolved residues and best 50% coverage template below 30% identity. b, Correlation between full chain TM-score and pTM on the same set (n = 151 chains), Pearson’s r = 0.84. The ground truth and predicted structure are shown for the most over-optimistic outlier (PDB: 6OFS, chain A). c, pTM distribution on a subset of the human proteome that we expect to be enriched for structurally novel multidomain proteins (n = 1,165 chains). Human proteome predictions comprise more than 600 confident residues (more than 50% coverage) and no proteins with 50% coverage templates. d, Four of the top hits from the set shown in c, filtering by pTM > 0.8 and sorting by number of confident residues. Proteins are labelled by their UniProt accession. For clarity, regions with pLDDT < 50 are hidden, as are isolated smaller regions that were left after this cropping.
Fig. 2 Full chain structure prediction. a, TM-score distribution for AlphaFold evaluated on a held-out set of template-filtered, long PDB chains (n = 151 chains). Includes recent PDB proteins with more than 800 resolved residues and best 50% coverage template below 30% identity. b, Correlation between full chain TM-score and pTM on the same set (n = 151 chains), Pearson’s r = 0.84. The ground truth and predicted structure are shown for the most over-optimistic outlier (PDB: 6OFS, chain A). c, pTM distribution on a subset of the human proteome that we expect to be enriched for structurally novel multidomain proteins (n = 1,165 chains). Human proteome predictions comprise more than 600 confident residues (more than 50% coverage) and no proteins with 50% coverage templates. d, Four of the top hits from the set shown in c, filtering by pTM > 0.8 and sorting by number of confident residues. Proteins are labelled by their UniProt accession. For clarity, regions with pLDDT < 50 are hidden, as are isolated smaller regions that were left after this cropping.CC BY · PubMed Central
Fig. 3 Highlighted structure predictions. a, Left, comparison of the active sites of two G6Pases (G6Pase-α and G6Pase-β) and a chloroperoxidase (PDB 1IDQ). The G6Pases are glucose-forming enzymes that contain a conserved, solvent-accessible glutamate (red; right) opposite the shared active-site residues (middle). b, Left, pocket prediction (P2Rank65) identifies a putative binding pocket for DGAT2, which is involved in body-fat synthesis. Red and green spheres represent the ligandability scores by P2Rank of 1 and 0, respectively. Middle, a proposed mechanism for DGAT151 activates the substrate with Glu416 and His415, which have analogous residues in the DGAT2 pocket. The docked inhibitor is well placed for polar interactions with His163 and Thr194 (right). The chemical structure (middle) is adapted from ref. 51. c, Predicted structure of wolframin, mutations in which cause Wolfram syndrome. Although there are regions in wolframin with low pLDDT (left), we could identify an OB-fold region (green/yellow), with a comparable core to a prototypical OB-fold (grey; middle). However, the most similar PDB chain (magenta; right) lacks the conserved cysteine-rich region (yellow) of our prediction. This region forms the characteristic β1 strand and an extended L12 loop, and is predicted to contain three disulfide bridges (yellow mesh).
Fig. 3 Highlighted structure predictions. a, Left, comparison of the active sites of two G6Pases (G6Pase-α and G6Pase-β) and a chloroperoxidase (PDB 1IDQ). The G6Pases are glucose-forming enzymes that contain a conserved, solvent-accessible glutamate (red; right) opposite the shared active-site residues (middle). b, Left, pocket prediction (P2Rank65) identifies a putative binding pocket for DGAT2, which is involved in body-fat synthesis. Red and green spheres represent the ligandability scores by P2Rank of 1 and 0, respectively. Middle, a proposed mechanism for DGAT151 activates the substrate with Glu416 and His415, which have analogous residues in the DGAT2 pocket. The docked inhibitor is well placed for polar interactions with His163 and Thr194 (right). The chemical structure (middle) is adapted from ref. 51. c, Predicted structure of wolframin, mutations in which cause Wolfram syndrome. Although there are regions in wolframin with low pLDDT (left), we could identify an OB-fold region (green/yellow), with a comparable core to a prototypical OB-fold (grey; middle). However, the most similar PDB chain (magenta; right) lacks the conserved cysteine-rich region (yellow) of our prediction. This region forms the characteristic β1 strand and an extended L12 loop, and is predicted to contain three disulfide bridges (yellow mesh).CC BY · PubMed Central
Fig. 4 Low-confidence regions. a, pLDDT distribution of the resolved parts of PDB sequences (n = 3,440,359 residues), the unresolved parts of PDB sequences (n = 589,079 residues) and the human proteome (n = 10,537,122 residues). b, Performance of pLDDT and the experimentally resolved head of AlphaFold as disorder predictors on the CAID Disprot-PDB benchmark dataset (n = 178,124 residues). c, An example low-confidence prediction aligned to the corresponding PDB submission (7KPX chain C)66. The globular domain is well-predicted but the extended interface exhibits low pLDDT and is incorrect apart from some of the secondary structure. a.a., amino acid. d, A high ratio of heterotypic contacts is associated with a lower AlphaFold accuracy on the recent PDB dataset, restricted to proteins with fewer than 40% of residues with template identity above 30% (n = 3,007 chains) (Methods). The ratio of heterotypic contacts is defined as: heterotypic/(intra-chain + homomeric + heterotypic).
Fig. 4 Low-confidence regions. a, pLDDT distribution of the resolved parts of PDB sequences (n = 3,440,359 residues), the unresolved parts of PDB sequences (n = 589,079 residues) and the human proteome (n = 10,537,122 residues). b, Performance of pLDDT and the experimentally resolved head of AlphaFold as disorder predictors on the CAID Disprot-PDB benchmark dataset (n = 178,124 residues). c, An example low-confidence prediction aligned to the corresponding PDB submission (7KPX chain C)66. The globular domain is well-predicted but the extended interface exhibits low pLDDT and is incorrect apart from some of the secondary structure. a.a., amino acid. d, A high ratio of heterotypic contacts is associated with a lower AlphaFold accuracy on the recent PDB dataset, restricted to proteins with fewer than 40% of residues with template identity above 30% (n = 3,007 chains) (Methods). The ratio of heterotypic contacts is defined as: heterotypic/(intra-chain + homomeric + heterotypic).CC BY · PubMed Central
Extended Data Fig. 1 Example full chain outputs containing both high- and low-confidence regions. Q06787 (synaptic functional regulator FMR1) and P54725 (UV excision repair protein RAD23 homologue A) are predicted to be disordered outside the experimentally determined regions by MobiDB73. Q92664 (transcription factor IIIA) has been described as ‘beads on a string’, consisting of zinc-finger domains joined by flexible linkers74.
Extended Data Fig. 1 Example full chain outputs containing both high- and low-confidence regions. Q06787 (synaptic functional regulator FMR1) and P54725 (UV excision repair protein RAD23 homologue A) are predicted to be disordered outside the experimentally determined regions by MobiDB73. Q92664 (transcription factor IIIA) has been described as ‘beads on a string’, consisting of zinc-finger domains joined by flexible linkers74.CC BY · PubMed Central
Extended Data Fig. 2 Distribution of per-residue lDDT-Cα within eight pLDDT bins. This represents an alternative visualization to Fig. 1a that does not sample the data. It uses the recent PDB dataset (Methods), which is restricted to structures with a reported resolution of <3.5 Å (n = 2,756,569 residues). Residues were assigned to bins of width 10 based on their pLDDT (minimum, 20; maximum, 100). Markers show the mean lDDT-Cα within each bin, while the lDDT-Cα distribution is visualized as a Matplotlib violin plot (kernel density estimate bandwidth, 0.2). The smallest sample size for the corresponding violin is 5,655 residues for the left-most bin.
Extended Data Fig. 2 Distribution of per-residue lDDT-Cα within eight pLDDT bins. This represents an alternative visualization to Fig. 1a that does not sample the data. It uses the recent PDB dataset (Methods), which is restricted to structures with a reported resolution of <3.5 Å (n = 2,756,569 residues). Residues were assigned to bins of width 10 based on their pLDDT (minimum, 20; maximum, 100). Markers show the mean lDDT-Cα within each bin, while the lDDT-Cα distribution is visualized as a Matplotlib violin plot (kernel density estimate bandwidth, 0.2). The smallest sample size for the corresponding violin is 5,655 residues for the left-most bin.CC BY · PubMed Central
Extended Data Fig. 3 Relationship between pLDDT and side-chain χ1 correctness. Evaluated on the recent PDB dataset (Methods), which is restricted to structures with a reported resolution of <2.5 Å (n = 5,983 chains) and residues with a B-factor of <30 Å2 (n = 609,623 residues). Residues are binned by pLDDT, with bin width 5 between 20 and 70 pLDDT and bin width 2 above 70 pLDDT. A χ1 angle is considered correct if it is within 40° of its value in the PDB structure75. Markers show the proportion of correct χ1 angles within each bin; error bars indicate the 95% confidence interval (two-sided Student’s t-test). The smallest sample size for the error bars is 193 residues for the left-most bin.
Extended Data Fig. 3 Relationship between pLDDT and side-chain χ1 correctness. Evaluated on the recent PDB dataset (Methods), which is restricted to structures with a reported resolution of <2.5 Å (n = 5,983 chains) and residues with a B-factor of <30 Å2 (n = 609,623 residues). Residues are binned by pLDDT, with bin width 5 between 20 and 70 pLDDT and bin width 2 above 70 pLDDT. A χ1 angle is considered correct if it is within 40° of its value in the PDB structure75. Markers show the proportion of correct χ1 angles within each bin; error bars indicate the 95% confidence interval (two-sided Student’s t-test). The smallest sample size for the error bars is 193 residues for the left-most bin.CC BY · PubMed Central
Extended Data Fig. 4 AlphaFold performance at a range of template sequence identities. lDDT-Cα for AlphaFold and BestSingleStructuralTemplate on 1 year of CAMEO targets39. Targets are binned according to the sequence identity of the best template covering at least 70% of the target, and a box plot is shown for each bin. The horizontal line indicates the median, boxes range from the lower to the upper quartile, and the whiskers extend from the minimum to the maximum. In total, 428 targets are included (see Source Data); the smallest number of targets in any bin is 18 Source data.
Extended Data Fig. 4 AlphaFold performance at a range of template sequence identities. lDDT-Cα for AlphaFold and BestSingleStructuralTemplate on 1 year of CAMEO targets39. Targets are binned according to the sequence identity of the best template covering at least 70% of the target, and a box plot is shown for each bin. The horizontal line indicates the median, boxes range from the lower to the upper quartile, and the whiskers extend from the minimum to the maximum. In total, 428 targets are included (see Source Data); the smallest number of targets in any bin is 18 Source data.CC BY · PubMed Central
Extended Data Fig. 5 Docking poses for a DGAT1-specific inhibitor in DGAT2. a, Top binding pose from Autodock Vina for a DGAT1-specific inhibitor in DGAT2, which does not match the predicted binding pocket for a DGAT2-specific inhibitor. b, Next best binding pose, which matches the binding pocket for the DGAT2-specific inhibitor, but does not contain components that satisfy the polar side chains His163 and Thr194. c, Relative positions of both binding poses.
Extended Data Fig. 5 Docking poses for a DGAT1-specific inhibitor in DGAT2. a, Top binding pose from Autodock Vina for a DGAT1-specific inhibitor in DGAT2, which does not match the predicted binding pocket for a DGAT2-specific inhibitor. b, Next best binding pose, which matches the binding pocket for the DGAT2-specific inhibitor, but does not contain components that satisfy the polar side chains His163 and Thr194. c, Relative positions of both binding poses.CC BY · PubMed Central
Extended Data Fig. 6 Relationship between sequence length and inference time. On the basis of logs from our human proteome set. All of the processed proteins are shown (n = 20,296). Each point indicates the mean inference time for the protein over the models produced. Vertical lines show the length cut-offs above which sequences were processed by multi-GPU workers.
Extended Data Fig. 6 Relationship between sequence length and inference time. On the basis of logs from our human proteome set. All of the processed proteins are shown (n = 20,296). Each point indicates the mean inference time for the protein over the models produced. Vertical lines show the length cut-offs above which sequences were processed by multi-GPU workers.CC BY · PubMed Central
Extended Data Fig. 7 Relationship between sequence length and run time for the non-inference stages of the pipeline. On the basis of 240 human protein sequences, chosen by stratified sampling from the length buckets: [16, 500), [500, 1,000), [1,000, 1,500), [1,500, 2,000), [2,000, 2,500) and [2,500, 2,700]. The relax plot shows five times more points, since five relaxed models are generated per protein. Coefficients for the quadratic lines of best fit were computed with Numpy polyfit.
Extended Data Fig. 7 Relationship between sequence length and run time for the non-inference stages of the pipeline. On the basis of 240 human protein sequences, chosen by stratified sampling from the length buckets: [16, 500), [500, 1,000), [1,000, 1,500), [1,500, 2,000), [2,000, 2,500) and [2,500, 2,700]. The relax plot shows five times more points, since five relaxed models are generated per protein. Coefficients for the quadratic lines of best fit were computed with Numpy polyfit.CC BY · PubMed Central

Verification

Claims the paper makes

14 Valid · 1 Not valid · 5 Not checked · 44 Own results and interpretations

36 more claims · 4 with their cited sources checked

Create a free account to see all 64 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 13
  • Probably not a substantial claim. 12

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

16 more claims about this paper

Create a free account to see every claim other papers make about this one.

See all claims — free account

Citations

Cites

Show all 83 references

Cited by

See all 3,049

Metadata

Authors
Tunyasuvunakool, K., Adler, J., Wu, Z., Green, T., Zielinski, M., Žídek, A., Bridgland, A., Cowie, A., Meyer, C., Laydon, A., Velankar, S., Kleywegt, G. J., Bateman, A., Evans, R., Pritzel, A., Figurnov, M., Ronneberger, O., Bates, R., Kohl, S. A. A., Potapenko, A., Ballard, A. J., Romera-Paredes, B., Nikolov, S., Jain, R., Clancy, E., Reiman, D., Petersen, S., Senior, A. W., Kavukcuoglu, K., Birney, E., Kohli, P., Jumper, J., & Hassabis, D.
Affiliations
  • Google DeepMind (United Kingdom) (GB)
  • European Bioinformatics Institute (GB)
Corresponding author
  • Tunyasuvunakool, Kathryn
  • Jumper, John
  • Hassabis, Demis
Venue
Nature
Published
2021-07-22
Volume
596
Issue
7873
Pages
590-596
ISSN
  • 0028-0836 (ISSN-L)
  • 1476-4687
OpenAlex
W3183475563
Cited by
3,025
References
83
Citations by year
2026: 346 · 2025: 517 · 2024: 580 · 2023: 701 · 2022: 692 · 2021: 181
FWCI
169.53
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature (publishedVersion, cc-by)
  • Europe PMC (PubMed Central) (publishedVersion, other-oa)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
Funders
  • European Bioinformatics Institute
  • DeepMind
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Protein Structure and Dynamics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Machine Learning in Bioinformatics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Structural Biology › Enzyme Structure and Function
Keywords
protein structure prediction, human proteome, AlphaFold2, structural coverage, intrinsically disordered regions, multidomain proteins, protein structure database
MeSH terms
Deep Learning, Glucose-6-Phosphatase, Humans, Membrane Proteins, Models, Molecular, Protein Conformation, Reproducibility of Results, Protein Folding, Computational Biology, Proteome, Diacylglycerol O-Acyltransferase, Datasets as Topic

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Highly accurate protein structure prediction for the human proteome. https://app.cochat.ai/papers/10.1038/s41586-021-03828-1
The paper
Tunyasuvunakool, K., Adler, J., Wu, Z., Green, T., Zielinski, M., Žídek, A., Bridgland, A., Cowie, A., Meyer, C., Laydon, A., Velankar, S., Kleywegt, G. J., Bateman, A., Evans, R., Pritzel, A., Figurnov, M., Ronneberger, O., Bates, R., Kohl, S. A. A., . . . Hassabis, D. (2021). Highly accurate protein structure prediction for the human proteome. Nature. https://doi.org/10.1038/s41586-021-03828-1

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.