Protein Data Bank

The Protein Data Bank (PDB) is the public archive of the three-dimensional shapes of proteins, nucleic acids and the complexes they form, as measured by experiment. It began in 1971 at Brookhaven National Laboratory with seven structures, and on October 25, 2026 it held 261,546, of which 208,325 were solved by X-ray diffraction, 37,405 by electron microscopy and 14,751 by nuclear magnetic resonance.^[1][2] Each entry is a table of atom coordinates under a four-character identifier, and journals that publish new structures expect them to be deposited here first.^[3]

Its entries trained AlphaFold. They also supply the monsters of the game Moleculoids.

The archive grew out of a symposium at Cold Spring Harbor in 1971, "Structure and Function of Proteins at the Three-Dimensional Level", and was set up at Brookhaven under Walter Hamilton.^[3][1] The journal Nature New Biology announced it on October 20, 1971.^[4][a] Before the Internet, the data moved by mail, "boxes of punched cards or magnetic tapes of coordinates".^[3] The file format dates to 1974 and keeps the fixed-width layout of that era, which limited the size of structures it could describe; the successor, mmCIF, was introduced in 1996.^[3]

Leadership passed from Hamilton to Tom Koetzle in 1973 and to Joel Sussman in 1994. In 1998 the management of the archive went to the Research Collaboratory for Structural Bioinformatics (RCSB), which brought together Rutgers, the University of California San Diego and the National Institute of Standards and Technology.^[1] A guideline effort of 1989 led by Fred Richards helped make deposition a condition of publication, and "virtually all journals that publish crystal structures of biomacromolecules" now require it.^[3] In 2003 the RCSB PDB, the Protein Data Bank in Europe and the Protein Data Bank Japan formed the Worldwide Protein Data Bank, so that there is "a single PDB archive" that stays freely available; electron microscopy and nuclear magnetic resonance data banks have since joined.^[1][5][6]

The archive reached 10,959 entries by the end of 1999 by its present count, and 100,000 in 2014.^[2][3] It added 17,601 in 2025.^[2]

X-ray crystallography still supplies about four entries in five. Electron microscopy has caught up in the newest ones: in 2015 the archive released 213 electron microscopy entries and 8,589 by X-ray, and in 2025 it released 7,198 and 10,082.^[2]

The structure-prediction programs learned from the archive. The paper on AlphaFold 3 states that "no structural data used during training were released after 30 September 2021", and by the archive's present count 182,455 entries had been released by that date.^[10][2] The paper also describes enriching the training data with structures predicted by AlphaFold-Multimer, so some of the structures the model learned from were predictions.^[10] A 2023 paper from Meta's research group FAIR applied a protein language model to more than 617 million metagenomic sequences, and predicted structures for all of them.^[11] The archive holds 261,546 measured ones.

Molecular Reality Corporation cites the archive in its argument about "the brain" and "the eye". The company's claim is that "AlphaFold 3 learned mostly from the half-century-old Protein Data Bank", and that the next models will need measurements that no archive holds yet; see Brain lab.^[12]

Its browser game Moleculoids draws its monsters from the archive. In May 2026 a script searched the RCSB PDB for entries in three ranges of molecular weight, under 100 kilodaltons, 100 to 500 and over 500, downloaded the coordinates and rendered 60 of them as sprites, which a review page, Gallerioid, lets a reviewer mark Good or Bad.^[12] The boss sprite is entry 11YE, a mouse Piezo1 ion channel in a membrane vesicle, which the archive released on April 29, 2026, the month before the script ran.^[12][13]

• AlphaFold
• Biological nanopore
• Brain lab
• Moleculoids