Read on Twitter
For some reason no one has yes posted on my feed a newslink to @DeepMind achievement in protein folding prediction. So I have. It’s ‘mostly’ very cool (see thread). @PhysicsofLifeUK @BritBiophysSoc @BiophysicalSoc https://apple.news/AXor5xubGTeSGgqhfGOZvww
This is a 50 year problem since Alfinsen showed in vitro that proteins had the information somehow in their amino acid sequence to determine their fold.
But. BUT. That fold is a needle in a hyper-dimensional haystack of such vastness that no random search could ever fold a protein in the age of the universe. So ...
The primary sequence must ALSO code for the Ariadne’s thread that leads into (in this case) the hyper dimensional labyrinth. That’s the hard bit.
I once tried a new approach to understanding the problem by dropping all the fancy potential stuff and keeping the high dimensional search, which everyone seemed to have forgotten but which seemed to me the most vital thing https://pure.york.ac.uk/portal/en/publications/protein-folding-in-highdimensional-spaces(a1e35af3-586a-4952-bd12-0d8e5874f9cb).html
Helped understand some of Sheena Radford’s folding mutants through the essential pathway contribution of non-native interactions.
So does the new AI approach help us understand how proteins fold?
Sadly not. That’s the problem with ‘AI’ - it’s a black box. It gives answers. Like ‘42’. But science is about questions.
Don’t get me wrong. It’s fantastic. We might be able to find protein structures for almost anything in silico now. Perhaps. But have we solved THE ‘protein folding problem’?
No
