Looking through the forum, I didn't see any mention of whether FoldIt was or could pull from the AlphaFold repo to provide any new tools or integrations.
Anyone planning to play around with ways to use it to improve FoldIt?
If I understand correctly, the output of the AlphaFold neural networks is predictions of the distances between amino acids in the protein. My guess is the pipeline that might lead to AlphaFold-style distance data eventually appearing in foldit is if someone first finds a useful way to combine it with Rosetta, which is the protein prediction software that foldit is an offshoot of. The scoring engine in foldit is a subset of the Rosetta software. Maybe distance data could be used as constraints for Rosetta, the same way contact data have been successfully used in the past.
The AlphaFold team actually used Rosetta in their pipeline for CASP (the big competition they won that made them famous). I'm really curious to know what role Rosetta played - did they use it for refinement only? For model validation or model picking? Or did they find a way to use their distance data in combination with the Rosetta engine to actually produce their models (I kinda doubt it was this last one, since they mention that their method did not involve complex sampling procedures, which I think are an integral part of Rosetta). I wonder if any of the foldit scientists know how Rosetta was used.
Here is a nice article about the AlphaFold project, including links to the Nature paper they produced and their open source code: https://techxplore.com/news/2020-01-alphafold-protein.html By the way, the little video at the top of that article is of CASP target T1008, which was a foldit player design (foldit3) and appeared in foldit's Nature paper. It's interesting that they chose it as an illustration, given that it does not have any sequence homolog data (because it is designed, not evolved) to feed into the neural network. I wonder what the input was for that one.
Sorry for the delayed reply! You're exactly right, the AlphaFold team did use Rosetta in two different steps of their CASP pipeline. However, it seems that Rosetta was not essential for AlphaFold's success, and was mostly helpful for refining the the AlphaFold models:
1. Preventing clashes: In combination with their distance predictions, AlphaFold used a coarse-grained version of the Rosetta score function (score2_smooth) to fold their models. They say this score function was mostly useful for preventing clashes. This makes sense: their neural net predicts distances between residues, but it doesn't know anything about the physical size of these residues, or how close the residues have to be in order to pack nicely.
2. Refinement: After AlphaFold has folded their model using the distance predictions and the coarse-grained Rosetta, they refine it using a Rosetta protocol called Relax (this protocol is very similar to the Foldit recipe BlueFuse).
Hoping to put some more info on the blog soon, about the potential impact of AlphaFold on Rosetta and Foldit. Stay tuned!
protein synthesis analysis, karelian