Jump to content
[dismiss]
Molecupedia is a parody encyclopedia written by Molecular Reality Corporation. It is not Wikipedia and is not affiliated with the Wikimedia Foundation. About

AlphaFold

Page protected
From Molecupedia, the encyclopedia of very small things
This is the current revision of this page. The present address (URL) is a permanent link to this version.
Artificial intelligence programs that predict protein structures
AlphaFold
Developer(s)Google DeepMind; AlphaFold 3 with Isomorphic Labs
Initial release2018 (CASP13)
Stable releaseAlphaFold 3 (May 2024; code November 2024)
Written inPython
TypeProtein structure prediction
LicenseApache 2.0 (code); trained parameters of AlphaFold 3 under separate terms of use
Websitealphafold.ebi.ac.uk

AlphaFold is a series of artificial intelligence programs, developed by Google DeepMind, that predict the three-dimensional structure of a protein from its sequence of amino acids. The second version, AlphaFold 2, reached a median score of 92.4 on the 100-point scale of the 2020 Critical Assessment of protein Structure Prediction (CASP14), a result the assessment's organizers recognized as a solution to a problem that had been open for about 50 years.[1] With the European Bioinformatics Institute, DeepMind has published more than 200 million predicted structures in a free database.[2] AlphaFold 3, released in 2024 with Isomorphic Labs, extends the predictions to complexes of proteins with DNA, RNA and small molecules.[3]

Half of the 2024 Nobel Prize in Chemistry went to Demis Hassabis and John Jumper of Google DeepMind "for protein structure prediction".[4] In the vocabulary of Molecular Reality Corporation, a nanopore company, AlphaFold is where "the brain" begins. Isomorphic Labs, the first name in the company's column of AI laboratories, was set up to carry AlphaFold into drug design, and the company cites AlphaFold as a model trained on data that earlier instruments collected.[5][6]

Background

A protein is a chain of amino acids that folds into a specific shape, and the shape determines what it does. Reading the amino acid sequence from DNA is cheap, while the shape has to be measured, usually by X-ray crystallography, nuclear magnetic resonance or cryo-electron microscopy, and a single structure can take months or years of work. The first structure of a protein, sperm whale myoglobin, was published in 1958.[7] The Protein Data Bank (PDB), the shared archive of measured structures, was established in 1971 at Brookhaven National Laboratory with seven entries.[8][9] In October 2026 it held 261,139 structures determined by experiment.[10]

CASP, held every two years since 1994, tests prediction methods blind: groups receive sequences of proteins whose structures have been measured but not yet released, and their predictions are scored against the experimental results. The main measure, the global distance test (GDT), runs from 0 to 100; a score of about 90 is informally considered competitive with experimental methods.[1]

Versions

AlphaFold (2018)

The first AlphaFold took part in CASP13 in 2018. It used a neural network to predict the distances between pairs of amino acids and then searched for a structure that satisfied them, and it was described in Nature in January 2020.[11]

AlphaFold 2 (2020)

AlphaFold 2 was entered in CASP14 in 2020 and achieved a median GDT of 92.4 across all targets, with an average error of about 1.6 ångströms, roughly the width of an atom.[1] The network, published in July 2021, reads an alignment of the protein's evolutionary relatives alongside a table describing every pair of residues, passes both through 48 blocks of a module called the Evoformer, and hands the result to a structure module that places the atoms.[12] It was trained on PDB structures released before April 30, 2018.[13] DeepMind released the code under the Apache 2.0 license the same month.[14]

On July 22, 2021, DeepMind and EMBL's European Bioinformatics Institute opened the AlphaFold Protein Structure Database with more than 350,000 predicted structures, including nearly the whole human proteome.[15][16] In July 2022 the database grew to 214,684,311 structures, covering almost every catalogued protein.[2][17] When the Nobel Prize was announced in October 2024, the Royal Swedish Academy of Sciences said AlphaFold 2 had been used by more than two million people in 190 countries.[4]

AlphaFold 3 (2024)

AlphaFold 3, described in Nature on May 8, 2024, was developed by Google DeepMind and Isomorphic Labs. It predicts the joint structure of complexes that can include proteins, DNA, RNA, small-molecule ligands, ions and chemically modified residues, using a diffusion model that generates atomic coordinates directly.[3] Researchers could first use it through a web service, the AlphaFold Server.[18] The code followed in November 2024; the trained parameters may be used only "if received directly from Google" and under separate terms of use.[18]

Confidence scores

The protein backbone is colored by a simulated per-residue confidence score (pLDDT), from blue (very high confidence) through yellow to orange (low confidence).
An animation from Molecular Reality's Iridian gallery colors a protein's backbone by a pLDDT score, from blue for very high confidence through yellow to orange. The scores in the animation are simulated.

Each prediction comes with two confidence scores. The predicted local distance difference test (pLDDT) is "a per-residue measure of local confidence", scaled from 0 to 100. Above 90 both the backbone and the side chains are usually right; below 50 the region is often disordered.[19] A second score, the predicted aligned error, estimates how well the relative positions of two parts of a structure are known. The database colors structures by pLDDT, dark blue for very high confidence through light blue and yellow to orange for very low.[19]

Limitations

AlphaFold predicts one structure for one sequence. In the European Bioinformatics Institute's training material, "A key limitation of protein structure prediction models is that they typically predict static structures as seen in the Protein Data Bank (PDB), and do not capture the dynamic behaviour of biomolecular systems in solution", and AlphaFold 3 "is no exception".[20] Because AlphaFold 3 is a generative model, it sometimes predicts spurious order, which the same material calls "hallucinations", in regions of a protein that are disordered.[20] The database's answers to frequently asked questions say that AlphaFold "has not been validated for predicting the effect of mutations", and a 2023 study that quoted them found very weak or no correlation between AlphaFold's output and measured changes in protein stability caused by single mutations.[21]

The input to AlphaFold is a sequence that someone has already read, and the output is a set of coordinates with confidence scores. The 214 million sequences behind the database came from UniProt, most of them translated from genomes read by DNA sequencers.[17] Which of those proteins a given drop of seawater contains, at what concentrations and in which modified forms, is a separate question, answered by instruments that examine the sample, such as mass spectrometers, sequencers and nanopores.

In Molecular Reality's usage

Main article: Brain lab

In the comparison that Molecular Reality draws between "the brain" and "the eye", Isomorphic Labs heads the brain column with the caption "AlphaFold + drug discovery AI".[5] Its materials state the dependence in one line: "AlphaFold 3 learned mostly from the half-century-old Protein Data Bank".[5] The PDB was filled over more than 50 years, one measured structure at a time, and the company argues that the next models will need new instruments to produce their data.[22]

The PDB appears elsewhere in the company's work. Its game Moleculoids draws its enemies from PDB structures.[23] The bacterial protein CsgG, from which Oxford Nanopore's sequencing pores were developed, entered the archive in September 2014 as entries 4UV2 and 4UV3, three and a half years before the cutoff date of AlphaFold 2's training data.[24][13]

See also

References

  1. ^ a b c "AlphaFold: a solution to a 50-year-old grand challenge in biology". Google DeepMind. November 30, 2020.
  2. ^ a b "AlphaFold predicts structure of almost every catalogued protein known to science". EMBL. July 28, 2022.
  3. ^ a b Abramson, Josh et al. (2024). "Accurate structure prediction of biomolecular interactions with AlphaFold 3". Nature 630: 493-500. doi:10.1038/s41586-024-07487-w.
  4. ^ a b "The Nobel Prize in Chemistry 2024" (Press release). Royal Swedish Academy of Sciences. October 9, 2024.
  5. ^ a b c "Company records". Molecular Reality Corporation. October 2026.
  6. ^ "Introducing Isomorphic Labs". Isomorphic Labs. November 2021.
  7. ^ Kendrew, J. C. et al. (1958). "A Three-Dimensional Model of the Myoglobin Molecule Obtained by X-Ray Analysis". Nature 181: 662-666. doi:10.1038/181662a0.
  8. ^ "History". RCSB Protein Data Bank. Retrieved October 25, 2026.
  9. ^ "Crystallography: Protein Data Bank". Nature New Biology 233: 223. 1971. doi:10.1038/newbio233223b0.
  10. ^ "RCSB Protein Data Bank". RCSB PDB. Retrieved October 25, 2026.
  11. ^ Senior, Andrew W. et al. (2020). "Improved protein structure prediction using potentials from deep learning". Nature 577: 706-710. doi:10.1038/s41586-019-1923-7.
  12. ^ Jumper, John et al. (2021). "Highly accurate protein structure prediction with AlphaFold". Nature 596: 583-589. doi:10.1038/s41586-021-03819-2.
  13. ^ a b "AlphaFold v2.3.0 technical note". GitHub. Google DeepMind.
  14. ^ "AlphaFold". GitHub. Google DeepMind. July 2021.
  15. ^ Varadi, Mihaly et al. (2022). "AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models". Nucleic Acids Research 50: D439-D444. doi:10.1093/nar/gkab1061.
  16. ^ Tunyasuvunakool, Kathryn et al. (2021). "Highly accurate protein structure prediction for the human proteome". Nature 596: 590-596. doi:10.1038/s41586-021-03828-1.
  17. ^ a b "New AlphaFold DB release with 200+ million predicted structures". EMBL-EBI. July 28, 2022.
  18. ^ a b "AlphaFold 3". GitHub. Google DeepMind. November 11, 2024.
  19. ^ a b "pLDDT: Understanding local confidence". EMBL-EBI Training. Retrieved October 25, 2026.
  20. ^ a b "What AlphaFold 3 struggles with". EMBL-EBI Training. Retrieved October 25, 2026.
  21. ^ Pak, Marina A. et al. (2023). "Using AlphaFold to predict the impact of single mutations on protein stability and function". PLOS ONE 18. doi:10.1371/journal.pone.0282689.
  22. ^ Dev Okonkwo-Hale (August 21, 2026). "The Startup That Says Everyone Else Is Building the Brain". The Picoampere.
  23. ^ Pixel Marchetti (July 14, 2026). "Moleculoids Review: An Asteroids Game That Turned Into a Dungeon". Ludicious.
  24. ^ Goyal, Parveen et al. (2014). "Structural and mechanistic insights into the bacterial amyloid secretion channel CsgG". Nature 516: 250-253. doi:10.1038/nature13768.
Retrieved from "https://molecularreality.com/wiki/AlphaFold"
Last edited 6 hours ago by DwellTimeDave