pocketz.doi.bio/2026-01-20_155501_nature_s41594-025-01725-z

The aging of the AlphaFold database | Nature Structural & Molecular Biology

URL: https://www.nature.com/articles/s41594-025-01725-z

Saved: 2026-01-20T23:55:06.729Z

Comment Published: 10 December 2025 The aging of the AlphaFold database Ifigenia Tsitsa, Anja Conev, Alessia David, Suhail A. Islam & Michael J. E. Sternberg

Nature Structural & Molecular Biology volume 32, pages 2374–2376 (2025)Cite this article

4074 Accesses

1 Citations

2 Altmetric

Metrics details

The AlphaFold database, released in 2022, modeled UniProt sequences from April 2021 and now provides 200 million predicted protein structures. Of the 20,504 full-length predicted human structures, 631 entries conflict with the June 2025 UniProt release. Similar conflicts across species highlight how bioinformatics resources can rapidly age.

You have full access to this article via your institution.

Download PDF

The AlphaFold database (AlphaFoldDB)1 is a major bioinformatics resource that provides predicted three-dimensional models for over 200 million protein sequences obtained from UniProt2. These models were generated using AlphaFold2 (ref. 3), the deep learning algorithm developed by DeepMind. These models are extensively used by the community for tasks including molecular replacement (for example, Phaser4 and MolRep5), missense variant prediction (for example, FoldX6 and Missense3D7) and as templates in structure prediction (for example, SwissModel8 and Phyre2.2 (ref. 9)). AlphaFoldDB was based on the protein sequences from April 2021 (UniProtKB version 2021_02, April 2021), which are outdated as UniProt sequences are regularly updated, a discrepancy that researchers should be aware of. Moreover, this is a major challenge for developers of general bioinformatics resources that require the maintenance of integrative and robust data management. Here, we have compared AlphaFold2 models with their respective UniProt sequences across six species and identified these discrepancies. It is vital for the community to be aware of these challenges to preserve the value of such resources.

In August 2025, we performed a comparison of AlphaFoldDB models with the latest UniProt (version 202503, that is, release 3 of 2025, published 18 June 2025) for the human proteome. We identified that among 20,504 unique UniProt accessions with AlphaFold2-predicted structures, 631 sequences differed, corresponding to a 3.08% discrepancy between AlphaFold2 and UniProt sequences (Fig. 1a). Out of the conflicting 631 models, 116 correspond to the UniProt sequence entries that were made redundant since the AlphaFoldDB construction that was based on UniProtKB release 202102 (ref. 10). The redundant UniProt sequences are generally the result of obsolete or reassigned genes, updated taxonomic data, changes in reference databases, annotation errors, or periodic data consolidation to improve consistency and accuracy (for example, UniProt ID Q5TG92). Sequence length mismatches were observed for 295 models, with UniProt sequences being longer than AlphaFoldDB for 155 models and shorter for 140 models. For example, ZNT2 (Q9BRI3), a zinc transporter protein implicated in the transient neonatal zinc deficiency, has a region of approximately 50 amino acids inserted in the new UniProt sequence around position 100 of the sequence of the outdated AlphaFoldDB model. To assess the effect of this insertion on the ZNT2 structure, we compared model structures (technical details in Supplementary Information). In particular, we compared the AlphaFoldDB model (Fig. 1b, blue) with an AlphaFold2 model generated from the updated UniProt sequence (Fig. 1b, green). In addition, we regenerated the AlphaFoldDB model using the same sequence and confirmed the low all-atom root mean square deviation (RMSD) between the AlphaFoldDB and our regenerated structure (Supplementary Fig. 1). We find that the insertion has major implications for the ZNT2 structure and locations of the transmembrane and zinc-binding domains (Fig. 1c,d). There were also 221 AlphaFoldDB models where the sequence length was the same, but the sequence differed. In the latter category is serpin B11 (Q96P15), a non-inhibitory intracellular serpin, with six amino acid substitutions between the two sequences.

Fig. 1: The discrepancy between AlphaFold2 models built on outdated and current UniProt sequences.

a, Sequence comparison of human AlphaFoldDB models with corresponding UniProt sequences. Blue rectangles indicate cases where the UniProt sequence length has changed, the red rectangle marks accessions that have become redundant, and green rectangles denote sequences where UniProt and AlphaFoldDB match in length. b, Structure models of proton-coupled zinc antiporter ZNT2 (Q9BRI3). The AlphaFold2-predicted structure downloaded from the AlphaFoldDB (blue) is shown aligned with a new AlphaFold2 model generated from the updated UniProt sequence (green). c, The transmembrane region is highlighted in orange in both the AlphaFoldDB model (blue) and the UniProt-updated model (green), showing the difference in localization of these regions. d, Zinc-binding residues are marked in red on both the AlphaFoldDB model (blue) and the UniProt-updated model (green), highlighting their conserved positions in the protein structure. Figures generated using PyMol (The PyMOL Molecular Graphics System, Version 3.1.0 Schrödinger, LLC).

Full size image

To assess the importance of proteins with differing sequences, we found that 406 out of the 631 had high annotation scores (4 and 5), 512 were ‘reviewed’ proteins and 136 were associated with diseases (as annotated by UniProt). Furthermore, 413 of these proteins were associated with Gene Ontology biological process terms and 418 with molecular function terms, which highlight their potential biological and functional relevance.

We expanded this analysis to include five additional model organisms (Table 1), with mouse showing a very similar trend to human, with 2.5% differing sequences. Interestingly, human and mouse sequences showed greater variation compared with other organisms, probably reflecting the increased depth of research and updates associated with these species. We further investigated this suggestion by looking at the ‘date of last sequence modification’ of UniProt entries across different species. We observed a consistent trend in a higher frequency of changes for human and mouse sequences since 2020 (Supplementary Fig. 2). On the other hand, Drosophila melanogaster and Arabidopsis thaliana had sequence differences of 0.32% and 0.18% respectively with Saccharomyces cerevisiae showing the least difference between AlphaFoldDB and UniProt sequences from all species (0.08%). Remarkably, Danio rerio shows a 46.68% difference in its sequences, as nearly half of its proteome is absent from the current UniProt version. This is owing to a major curation effort that removed thousands of obsolete or redundant entries, reducing the reviewed D. rerio proteome to 3,355 proteins, of which 93 are not present in the AlphaFoldDB. Notably, some entries, such as Q802X2 and B7ZC96, have AlphaFoldDB models but are excluded from the downloadable version of AlphaFoldDB, probably due to their annotation as ‘unreviewed’ structures. Reviewing their UniProt history reveals that they were only recently marked as reviewed (Q802X2 in July 2024 and B7ZC96 in June 2023), well after the last update of AlphaFoldDB in 2022, illustrating both sequence and annotation aging in the AlphaFoldDB. Excluding all the AlphaFoldDB models that no longer correspond to the updated UniProt database, the total number of differing protein sequences is reduced to 10, resulting in a UniProt AlphaFoldDB difference that aligns with trends observed in other organisms.

Table 1 Summary of the sequence comparison between UniProt and AlphaFoldDB Full size table

In conclusion, the AlphaFoldDB provides an unprecedented resource for structural biology, but its reliance on outdated sequence and annotation data poses challenges across its widespread applications, including structure-based drug design, variant interpretation and functional analysis. These discrepancies highlight the need for regular updates to maintain alignment with UniProt, especially as other groups, such as SWISS-MODEL, continue to develop their own structure databases. One way to mitigate these issues is by using UniProt as the primary reference and mapping sequences to the corresponding AlphaFold2 models through alignment. We stress the importance of routinely checking consistency between UniProt sequences and AlphaFold2 models, as underscored by UniProt’s and SWISS-MODEL’s new warning flags and recent community efforts to remodel outdated entries with AlphaSync (available at https://alphasync.stjude.org/)11. While finalizing this Comment, EMBL-EBI (European Molecular Biology Laboratory’s European Bioinformatics Institute) announced release of an updated version of the AlphaFoldDB that corresponds to the latest UniProt release (2025_03). However other resources that build on earlier deposited AlphaFold2 models will still be out of date and may not have implemented a warning if their structural model corresponds to an out-of-date UniProt entry. Addressing these issues will be crucial for maximizing the utility of AlphaFold2 models and ensuring their integration into downstream tools and resources. Similarly, static bioinformatics databases, including AlphaMissense12, which present precomputed missense variant effect predictions, are also aging. Time flies in bioinformatics.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability

All data used in this analysis is publicly available. The UniProtKB data (release 202503) is available at https://ftp.uniprot.org/pub/databases/uniprot/previousreleases/release-2025_03/. The AlphaFoldDB proteomes are available for download at https://ftp.ebi.ac.uk/pub/databases/alphafold/v4/.

Code availability

The analysis is performed using Bash and Python programming languages and the code is available in the Code Ocean capsule: https://codeocean.com/capsule/3156855.

References

Varadi, M. et al. Nucleic Acids Res. 50, D439–D444 (2022).

Article CAS PubMed Google Scholar

The UniProt Consortium Nucleic Acids Res. 51, D523–D531 (2023).

Article Google Scholar

Jumper, J. et al. Nature 596, 583–589 (2021).