Biology · Structural Biology

Protein Structures: Cumulative

Fifty years of lab work built 185,392 structures; one AlphaFold release added 214 million.

• 185,392 experimental structures in 50 years
• 214M predicted in one July 2022 release
• Portal says 261M; DeepMind’s share is 241M
• AlphaFold 2, not AlphaFold 3

This count of cumulative known protein structures changes basis partway: experimental entries from the RCSB Protein Data Bank through 2021, then those entries plus AI predictions from 2022. The two bases are kept distinct rather than blended, because a predicted structure is not a solved one. Structural biology spent fifty years building the experimental archive by X-ray crystallography, NMR and later cryo-EM, reaching 185,392 entries by the end of 2021, the count the archive published at the time. Then in July 2022 AlphaFold DB went, in EMBL-EBI’s own words, from less than a million to over 214 million structures: more than a thousand times the entire experimental record and covering nearly every protein sequence then catalogued. Those predictions are AlphaFold 2’s, not AlphaFold 3’s. AlphaFold 3 arrived in 2024 and extended prediction to DNA, RNA and small-molecule complexes, but the bulk database is AlphaFold 2 output. The database is not frozen. Its FTP changelog records a jump to 241,070,489 structures in September 2025, when it was re-synchronised to the UniProt 2025_03 reference. The experimental archive kept going, adding 15,457 structures in 2024 and 17,608 in 2025, and its total stands at 259,412. Those two annual figures are RCSB’s own count of entries first released in each of those years; they and the total were read on 4 September 2026. Subtracting the plotted points instead gives a dozen more for 2025. The plotted values are snapshots taken at the time, and the archive loses entries to obsoletion afterwards, so the two bases never land on exactly the same number. The 2026 change was to the database’s scope rather than AlphaFold’s output. EMBL-EBI began hosting sets computed by other groups running AlphaFold 2 through ColabFold: 17.7 million bacterial and archaeal proteins from AllTheBacteria, smaller viral and kinetoplastid sets and about 2 million predicted dimers contributed by NVIDIA. That pushes the portal’s own count past 261 million, but Google DeepMind’s contribution is still exactly 241,070,489, and that is the figure plotted here. The contributed sets come mostly from genome assemblies rather than UniProt and split long proteins into overlapping fragments, so they are not the same counting unit. The dimers are excluded as well, because a complex is a different object from a single-chain structure and is assembled from proteins already counted once.

Source: RCSB Protein Data Bank / EMBL-EBI AlphaFold DB

Download the data: alphafold.csv · alphafold.json

All curves · The chart is drawn in your browser and needs JavaScript. Every number it plots is in the table above.