Read on Twitter
Many people seem excited with AlphaFold2. The progress is impressive and surely has implications to many fields. Predicting structures of protein complexes and proteins that adopt multiple states (e.g. transporters) should be also within reach, by extension of the current method.
There are already papers that show MSA encode such information as well.
However impressive, though, I don't consider the protein folding problem "has been solved". I am not comfortable with reliance on multiple sequence alignment (MSA).
However impressive, though, I don't consider the protein folding problem "has been solved". I am not comfortable with reliance on multiple sequence alignment (MSA).
Yes, MSA provides a big hint, but E. coli can fold an exogenous sequence without "seeing" MSA. In other words, a single sequence contains enough information. If we really understand the physics of protein folding, we should be able to do the same in silico.
This has practical implications. Can AlphaFold2 estimate structures of highly conserved proteins with little variations within MSA? How about de novo designed proteins without any MSA?
Can it predict effects of mutations? Sequences in genomes have been selected by evolution. They encode information on what is necessary to keep a protein folded and functional, but probably not much on mutations that destabilize proteins to take non-native shapes (e.g. amyloids).
Other complimentary approaches from machine learning to protein folding include machine learning based force fields (cf. Rosetta uses hand-designed energy functions) and accelerated sampling and exploration of the energy landscape. I hope many people work on such approaches.
Some said they consider the problem has been solved and will leave the field; I feel opposite. Previously, we didn't even know if the problem is solvable, or how many years and computational cost it would take. Now we know it is tractable. I hope more people start working on this
