AlphaFold statistics [2026]

AlphaFold statistics are easy to misquote because "AlphaFold" can mean the model, the AlphaFold Protein Structure Database, AlphaFold Server, or newer complex-prediction datasets. The headline database number is now 241,070,489 predicted protein structures in AlphaFold DB v6.

This page separates the current database counts from model accuracy, AlphaFold 3 capabilities, complex predictions, and research adoption.

AlphaFold at a glance

The most useful AlphaFold statistics depend on the counting level. A database entry, a model run, a protein complex, a paper citation, and an experimental PDB structure all measure different things.

The bullets below are the headline figures worth keeping separate when citing AlphaFold in 2026.

• AlphaFold DB v6 contains 241,070,489 predicted protein structures, including 40,054 isoforms, in the October 21, 2025 release synced to UniProt 2025_03.^[1]
• Calculated: AlphaFold DB v6 is about 941 times the size of the experimental PDB archive, because 241,070,489 AlphaFold DB v6 predictions divided by 256,292 RCSB PDB archive structures on July 1, 2026 equals 940.6.^[1][10]
• The AlphaFold Database had more than 3.4 million users from 190 countries as of EMBL's March 16, 2026 complex-dataset announcement.^[5]
• In CASP14, AlphaFold 2 achieved 0.96 angstrom median backbone accuracy at 95% residue coverage, compared with 2.8 angstroms for the next best method.^[3]
• AlphaFold 3 was published in Nature in 2024 as a model for complexes containing proteins, nucleic acids, small molecules, ions, and modified residues.^[4]
• The 2024 Nobel Prize in Chemistry awarded one half jointly to Demis Hassabis and John M. Jumper "for protein structure prediction" and one half to David Baker "for computational protein design".^[9]

AlphaFold Database size and release history

AlphaFold DB is a moving resource. Its count changes when UniProt records are added, removed, updated, or represented by isoforms.

The current v6 release should not be mixed with older "over 200 million" or "over 214 million" figures unless the release date is named.

• AlphaFold DB began data archiving in July 2021 with more than 360,000 structures for 20 model-organism proteomes from UniProt release 2021_02.^[2]
• By September 2023, AlphaFold DB had expanded to more than 214 million predicted structures, according to the 2024 AlphaFold DB paper.^[2]
• AlphaFold DB v4 contained exactly 214,683,829 predicted protein structures, according to the v6 release notes.^[1]
• AlphaFold DB v6 contains exactly 241,070,489 predicted protein structures, including 40,054 isoforms.^[1]
• Calculated: AlphaFold DB v6 is 26,386,660 predictions larger than v4, because 241,070,489 minus 214,683,829 equals 26,386,660.^[1]
• Calculated: AlphaFold DB v6 is 12.3% larger than v4, because 26,386,660 additional predictions divided by 214,683,829 v4 predictions equals 0.1229.^[1]
• Between AlphaFold DB v4 and v6, 65,711,653 proteins were added, 39,324,993 were removed, 276,539 were changed, and 175,082,297 were unchanged but relabelled with current metadata.^[1]
• RCSB.org listed 1,062,058 computed structure models on July 1, 2026, a separate search set that includes computed models from AlphaFold DB and ModelArchive rather than the full AlphaFold DB corpus.^[10]

Coverage, files, and confidence scores

AlphaFold DB is not just a web search box. It is also a downloadable model archive with model files, metadata, and confidence measures that users need to interpret predictions correctly.

Confidence statistics are especially important because AlphaFold entries include coordinates even for regions that the model marks as uncertain.

• AlphaFold DB predictions are stored in PDB, mmCIF, and binaryCIF formats, with corresponding metadata in JSON format.^[2]
• The full AlphaFold DB v4 dataset was approximately 23 TiB on Google Cloud Public Datasets under a CC-BY-4.0 license.^[2]
• The 2024 AlphaFold DB paper estimated that downloading the complete v4 dataset would take roughly 2.5 days over a 1 Gbps internet connection.^[2]
• As of September 2023, the EMBL-EBI FTP area hosted versioned TAR files for 48 model organisms and pathogens, plus metadata files for AlphaFold DB.^[2]
• AlphaFold reports pLDDT on a 0 to 100 scale, and the AlphaFold DB paper describes pLDDT greater than 90 as generally suitable for high-accuracy uses such as binding-site characterization.^[2]
• AlphaFold DB guidance treats pLDDT 70 to 90 as a generally reliable backbone prediction, 50 to 70 as lower confidence, and below 50 as often indicating probable disorder.^[2]
• In the AFDB clustering analysis, 214 million UniProtKB protein sequences were reduced to 52 million sequence clusters, then 18.8 million structural clusters, then 2.30 million robust clusters with at least two structures.^[2]

Accuracy and benchmark statistics

Accuracy numbers should name the benchmark and metric. AlphaFold's strongest single-chain results do not automatically transfer to every region, interface, ligand, or flexible state.

The most cited AlphaFold 2 benchmark figures come from CASP14 and from a later evaluation on recently released PDB structures.

• The CASP14 AlphaFold 2 analysis used 87 protein domains for the main backbone-accuracy comparison.^[3]
• In CASP14, AlphaFold 2 reached 0.96 angstrom median C-alpha RMSD at 95% coverage, while the next best method reached 2.8 angstroms.^[3]
• In the same CASP14 analysis, AlphaFold 2 reached 1.5 angstrom all-atom RMSD at 95% coverage, compared with 3.5 angstroms for the best alternative method.^[3]
• On a filtered recent-PDB test set excluding close training templates, AlphaFold 2 had a 1.46 angstrom median backbone RMSD at 95% coverage across 3,144 protein chains.^[3]
• AlphaFold 2's pLDDT confidence score correlated with true lDDT-C-alpha accuracy at Pearson's r = 0.76 across 10,795 protein chains in the recent-PDB analysis.^[3]
• AlphaFold 2's pTM confidence estimate correlated with full-chain TM-score at Pearson's r = 0.85 across 10,795 protein chains in the recent-PDB analysis.^[3]
• Without ensembling, AlphaFold 2 representative neural-network inference times were 0.6 minutes for 256 residues, 1.1 minutes for 384 residues, and 2.1 hours for 2,500 residues on the reported setup.^[3]
• In the AlphaFold 2 paper, a 16 GB V100 GPU could predict proteins up to about 1,300 residues without ensembling, while the 2,500-residue case used unified memory and requested four GPUs for sufficient memory.^[3]

AlphaFold 3 and complex prediction

AlphaFold 3 changes the counting problem because it predicts biomolecular complexes, not only single protein chains. The useful statistics are therefore about molecular coverage, sampling, and complex datasets.

The 2026 complex-dataset updates are separate from AlphaFold DB's core v6 single-protein count.

• AlphaFold 3 was published online on May 8, 2024, with the version of record dated June 11, 2024 in Nature volume 630, pages 493-500.^[4]
• AlphaFold 3's paper describes a unified model for complexes containing proteins, DNA, RNA, small molecules, ions, and modified residues.^[4]
• AlphaFold 3 benchmark results in the Nature paper generally selected the top confidence-ranked result from 5 model seeds with 5 diffusion samples per seed, for 25 candidate samples.^[4]
• AlphaFold 3's pairformer module uses 48 blocks, while the MSA embedding block is reduced to 4 blocks compared with AlphaFold 2's Evoformer-heavy design.^[4]
• In March 2026, the AlphaFold Database complex collaboration reported that it had calculated predictions for 30 million complexes, added 1.7 million high-confidence homodimer predictions to the database, and made 18 million lower-confidence homodimers available for bulk download.^[5]
• In a May 19, 2026 update, EMBL reported that almost 80,000 high-confidence heterodimer predictions had been added to AlphaFold DB and 8.1 million lower-confidence heterodimer predictions were available for bulk download.^[5]
• The 2026 complex dataset would require around 17 million GPU hours to recreate, according to EMBL's March 2026 release.^[5]
• The 2026 AlphaFold complex update focuses on 20 of the most studied species, including humans, plus the World Health Organization priority pathogens list.^[5]

Community datasets and scientific adoption

AlphaFold's impact is partly a protein structure prediction story and partly an adoption story. The strongest adoption numbers come from first-party database reporting and an independent Innovation Growth Lab analysis.

Community datasets also matter because they extend AlphaFold DB beyond the better-covered model-organism and Swiss-Prot space.

• In February 2026, AlphaFold DB added more than 17 million bacterial protein structure predictions from the AllTheBacteria project as part of its community-dataset program.^[6]
• In February 2026, AlphaFold DB added more than 350,000 viral protein structure predictions from the Big Fantastic Virus Database.^[6]
• In February 2026, AlphaFold DB added Viro3D predictions covering more than 85,000 proteins from more than 4,400 viruses.^[6]
• Google DeepMind reported in late 2025 that AlphaFold had been used by more than 3 million researchers in more than 190 countries, including more than 1 million users in low- and middle-income countries.^[7]
• Google DeepMind reported that AlphaFold Server had produced more than 8 million folds for thousands of researchers by late 2025.^[7]
• The Innovation Growth Lab 2025 analysis studied 5 million academic publications, clinical articles, patents, and protein structures to estimate AlphaFold 2's scientific impact.^[8]
• The Innovation Growth Lab estimated that 550,000 publications were linked to AlphaFold 2 directly or indirectly, involving almost 2 million unique researchers, and that 218,000 articles incorporated AlphaFold 2 elements into their methodology.^[8]
• Researchers building on AlphaFold 2 showed a 45% to 49% increased rate of experimental protein-structure submissions in the Innovation Growth Lab analysis.^[8]
• Publications building on AlphaFold 2 had a 28.9% citation-count increase in the Innovation Growth Lab analysis.^[8]
• Papers linked to AlphaFold 2 were twice as likely to be cited in clinical articles, and AlphaFold 2-linked papers, researchers, and laboratories were 22.6% to 34.2% more likely to be cited by patents in the Innovation Growth Lab analysis.^[8]

Methodology and citation

This article uses primary or official sources where available: AlphaFold DB release notes, peer-reviewed AlphaFold papers, EMBL-EBI and EMBL database announcements, NobelPrize.org, RCSB PDB, Google DeepMind first-party impact reporting, and the Innovation Growth Lab impact report.

Calculated values are labeled in the bullet where they appear. Because AlphaFold DB, RCSB PDB, and usage metrics update over time, cite the date attached to each source-backed number.

• The v6 minus v4 calculation uses 241,070,489 minus 214,683,829 from the PDBe release notes, yielding 26,386,660 additional predictions.^[1]
• The 12.3% v6 growth calculation divides 26,386,660 additional predictions by 214,683,829 v4 predictions from the PDBe release notes.^[1]
• The 941-times PDB comparison divides 241,070,489 AlphaFold DB v6 predictions from the PDBe release notes by 256,292 experimental PDB archive structures shown on the RCSB PDB homepage on July 1, 2026.^[1][10]
• Suggested citation: ProteinIQ. "AlphaFold statistics [2026]." Updated July 1, 2026. Accessed [your access date]. https://proteiniq.io/guides/alphafold-statistics

Sources▼

1. AlphaFold Database release notes Protein Data Bank in Europe · July 1, 2026. https://www.ebi.ac.uk/pdbe/news/alphafold-database-release-notes
2. AlphaFold Protein Structure Database in 2024: providing structure coverage for over 214 million protein sequences Nucleic Acids Research · 2024. https://academic.oup.com/nar/article/52/D1/D368/7337620
3. Highly accurate protein structure prediction with AlphaFold Nature · 2021. https://pmc.ncbi.nlm.nih.gov/articles/PMC8371605/
4. Accurate structure prediction of biomolecular interactions with AlphaFold 3 Nature · 2024. https://www.nature.com/articles/s41586-024-07487-w
5. Millions of protein complexes added to AlphaFold Database shed light on how proteins interact EMBL · July 1, 2026. https://www.embl.org/news/science-technology/first-complexes-alphafold-database/
6. AlphaFold Database welcomes community datasets EMBL-EBI · July 1, 2026. https://www.ebi.ac.uk/about/news/updates-from-data-resources/alphafold-database-community-datasets/
7. AlphaFold: Five Years of Impact Google DeepMind · July 1, 2026. https://deepmind.google/blog/alphafold-five-years-of-impact/
8. AI in Science: Evidence of impact from AlphaFold 2 Innovation Growth Lab · 2025. https://www.innovationgrowthlab.org/wp-content/uploads/2025/11/ai_in_science_af2_igl_summary.pdf
9. Press release: The Nobel Prize in Chemistry 2024 NobelPrize.org · 2024. https://www.nobelprize.org/prizes/chemistry/2024/press-release/
10. RCSB PDB homepage counts RCSB PDB · July 1, 2026. https://www.rcsb.org/