Nobel Chemistry Prize goes to AlphaFold, Rosetta creators - another win for AI
- Reference: 1728498565
- News link: https://www.theregister.co.uk/2024/10/09/alphafold_rosetta_nobel_chemistry_prize/
- Source link:
DeepMind cofounder and CEO Demis Hassabis and director John Jumper will share half of the Nobel Prize in Chemistry for their work on AlphaFold models. The second generation can predict almost all known protein structures - [1]more than 200 million in total.
"The team trained AlphaFold2 on the vast information in the databases of all known protein structures and amino acid sequences and the new AI architecture started delivering good results," the Nobel committee [2]said [PDF].
[3]
When it entered the 2020 Critical Assessment of Protein Structure Prediction (CASP) competition, AlphaFold2 performed almost as well as X-ray crystallography (the prior gold standard in modeling protein structures) "in most cases," the committee added. "Previously, it often took years to obtain a protein structure, if at all. Now it can be done in a few minutes."
[4]
[5]
Jumper, who came to DeepMind after the Google subsidiary had already built the initial AlphaFold that improved on prior CASP results but was still only about 60 percent accurate, was essential to DeepMind 2's success, the Nobel body said.
"AlphaFold2 was coloured by Jumper's knowledge of proteins," the committee explained. "The team also started to use the innovation behind the recent enormous breakthrough in AI: neural networks called transformers."
[6]
So maybe some additional AI tech helped, too.
Your bespoke protein is ready
Although AlphaFold has been fundamental in helping humans become better predictors of protein shapes, which play a critical role in their function, it can't develop drugs or make anything new.
That's where Rosetta, designed by University of Washington biochemistry professor David Baker, comes in.
Baker developed his own protein prediction software, dubbed Rosetta, in the 1990s, and when it entered the CASP competition in 1998, it did well "in comparison to other participants," the Nobel committee said. After the competition, Baker and his team got the idea to use the software in reverse: Instead of using amino acid sequences to predict the shape of a protein, they began experimenting on whether inputting the shape of a desired protein would suggest an amino acid sequence to create it.
[7]AI and robots join forces to cook up proteins faster
[8]DeepMind spinoff Isomorphic claims AlphaFold 3 predicts bio-matter down to the DNA
[9]AI drug algorithms can be flipped to invent bioweapons
[10]Boffins deem Google DeepMind's material discoveries rather shallow
Lo and behold, it worked perfectly and led to the creation of Top7, "the first protein that was entirely different to all other known existing proteins," according to the Nobel folks.
Proteins are fundamental to understanding biochemistry and are involved in the creation of biological structures like muscles, as well as chemicals like hormones and antibodies. By enabling the creation of new proteins, humans can do all sorts of things.
"This can lead to new nanomaterials, targeted pharmaceuticals, more rapid development of vaccines, minimal sensors and a greener chemical industry – to name just a few applications," the committee said.
[11]
The chemistry Nobel being awarded for AI development marks the second time this year. The Nobel in physics was [12]awarded yesterday to John Hopfield for his work developing early neural networks, and to AI godfather Geoffrey Hinton for giving machines the ability to interpret information they're trained to recall.
Three Nobels have been awarded so far this year; the first, for physiology and medicine, went to Victor Ambros and Gary Ruvkun for the [13]discovery of microRNA, which regulates gene expression and protein production. Nobel prizes for literature and peace have not yet been handed out. ®
Get our [14]Tech Resources
[1] https://www.theregister.com/2022/07/28/deepmind_alphafold_protein_folding/
[2] https://www.nobelprize.org/uploads/2024/10/popular-chemistryprize2024-3.pdf
[3] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_offbeat/science&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2Zwb9BtFJjItPH3TcefCVGwAAAMY&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_offbeat/science&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Zwb9BtFJjItPH3TcefCVGwAAAMY&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_offbeat/science&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Zwb9BtFJjItPH3TcefCVGwAAAMY&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[6] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_offbeat/science&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Zwb9BtFJjItPH3TcefCVGwAAAMY&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[7] https://www.theregister.com/2024/01/15/ai_robot_protein_engineering/
[8] https://www.theregister.com/2024/05/09/google_deepmind_alphafold3_model/
[9] https://www.theregister.com/2022/03/18/ai_weapons_learning/
[10] https://www.theregister.com/2024/04/11/google_deepmind_material_study/
[11] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_offbeat/science&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Zwb9BtFJjItPH3TcefCVGwAAAMY&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[12] https://www.theregister.com/2024/10/08/ai_godfather_wins_nobel_prize/
[13] https://www.nobelprize.org/prizes/medicine/2024/press-release/
[14] https://whitepapers.theregister.com/
Indeed. And it would be more worthy of a prize IMHO if it could actually identify unknown new useful proteins, rather than " almost all known protein structures " -- which we already know.
I think we might have finally reached [1]The Singularity here, the localized inversion layer at which [2]Ig Nobel Prizes suddenly become much more relevant than Nobel ones! (on top of being more entertaining!)
[1] https://en.wikipedia.org/wiki/The_singularity
[2] https://www.theregister.com/2024/09/14/ig_nobel_prize_2024/
'rather than "almost all known protein structures" -- which we already know'
Umm, no. We don't. The very next bit of the article is " - more than 200 million in total." and we certainly don't know 200 million protein structures. As I understand it, in the early days people used to sanity-check AlphaFold's predictions quite thoroughly. Perhaps not spending "years" on X-ray crystallography but certainly some substantial time. As the community slowly got used to the idea that "damn, the blasted computer is always right", the checking has become less and less. Nowadays, it is often omitted because frankly researchers can't justify spending time or money on something that is very unlikely to produce a surprise. Also, anyone who does check the result is going to be beaten by others to the interesting consequences.
In short, then, AlphaFold and its ilk have transformed this area of biology by being a magic box that produces protein structures in minutes which are as accurate as the ones you could produce in several years by traditional methods (and even that's assuming the protein actually exists in Nature to be X-rayed).
Odd that they've got the Chemistry prize for it though, since the (sole?) application is clearly to Biology and Medicine.
I suppose it's more chemistry than the physics prize was actually physics, but still I think this tarnishes the Nobel committee for chasing bandwagons and finding excuses to give awards to an area of computer science.