For half a century, biologists solved protein structures one at a time, with X-ray crystallography, NMR and cryo-electron microscopy. Each structure took months to years of lab work. The Protein Data Bank, the public archive for all of them, holds 260,724 entries.

In July 2021, DeepMind published AlphaFold 2. The AlphaFold Protein Structure Database, run with EMBL-EBI, now holds more than 200 million predicted structures. That is more than 750 times the number of structures science solved by experiment, from a single method.

How big is the gap?

Source: RCSB Protein Data Bank search API, October 4, 2026; AlphaFold Protein Structure Database, EMBL-EBI. Predictions and experimental structures are not the same kind of evidence. See the limits below.

View data

archive	structures
AlphaFold DB (predicted)	200000000
Protein Data Bank (experimental)	260724

The comparison is lopsided because the two archives grow by different mechanisms. The Protein Data Bank grows when a lab solves a structure and deposits it. The AlphaFold database grew when a model was run across protein sequences at scale.

Did AI replace the experiments?

No. The same archive tells the opposite story.

1. 12020: AlphaFold 2 at CASP14
2. 22021: AlphaFold 2 paper and database launch

Source: RCSB Protein Data Bank search API, by initial release date, October 4, 2026. 2026 is excluded because the year is incomplete (13,499 released through October 4).

View data

year	entries
2000	2624
2001	2810
2002	2994
2003	4144
2004	5149
2005	5334
2006	6405
2007	7134
2008	6902
2009	7288
2010	7742
2011	7924
2012	8750
2013	9338
2014	9551
2015	9242
2016	10803
2017	11061
2018	11169
2019	11479
2020	14000
2021	12584
2022	14284
2023	14498
2024	15455
2025	17602

Releases rose from 2,624 in 2000 to 17,602 in 2025, and every year since 2020 has been higher than the one before except 2021. The three years after AlphaFold 2 (2022 to 2024) were all above 14,000. Prediction did not displace experiment. The two are growing side by side.

Why might they be growing together?

The data cannot tell us why, so this part is interpretation. A predicted structure is a hypothesis about shape. Labs still need to confirm that the shape is right, and they still need experiments for what predictions handle poorly: complexes, ligands bound to a drug, and the effects of mutations. A free database of 200 million hypotheses may increase the demand for experimental checks, not reduce it.

The Nobel Prize in Chemistry in 2024 reflected both halves of the field. Demis Hassabis and John Jumper were recognized for protein structure prediction, and David Baker for computational protein design.

Limits

A predicted structure is not a measured one. AlphaFold reports a confidence score for each part of a prediction, and low-confidence regions should not be treated as structures. We counted entries, not quality, and the count of 200 million is the lower bound the database states, not an exact figure. The Protein Data Bank also contains many entries for the same protein in different states, so 260,724 entries is not 260,724 different proteins.

Sources

1. RCSB Protein Data Bank, search API: entry counts by initial release date, retrieved October 4, 2026. Total entries released through 2026: 260,724.
2. AlphaFold Protein Structure Database, EMBL-EBI: "over 200 million" predicted structures.
3. Jumper et al., Highly accurate protein structure prediction with AlphaFold, Nature, 2021.
4. Nobel Prize, The Nobel Prize in Chemistry 2024.