Abstract
Protein folding has been a “holy grail” problem of biology for fifty years, and the recent breakthrough from AlphaFold2 and RoseTTAfold set a profound milestone for solving this problem. AlphaFold2 and RoseTTAfold were successful in predicting protein structures from peptide sequences with high accuracy. Meanwhile, although the protein folding problem also cares about the kinetic pathways of protein folding, AlphaFold2 and RoseTTAfold were not trained with this functionality. Considering that their training sets contain proteins from foldable sequences and sequence evolutionary information, we wondered if the computational models from AlphaFold2 and RoseTTAfold might carry protein foldability information. To test this idea, we systematically predicted the structural models of 149 circular permutants and 148 alanine insertion mutants of the 149-residue dihydrofolate reductase of Escherichia coli with AlphaFold2 and RoseTTAfold. Our data showed that although AlphaFold2 and RoseTTAfold could not directly identify unfoldable proteins, the structural variations of computational models are correlated with protein foldability. Furthermore, this correlation is independent of secondary structures. Most importantly, the structural variations of computational models are quantitatively correlated with protein foldability but not protein function. Our work could be of great value to the design of circular permutants, the design of fragment complementary proteins, the design of novel proteins, and the development of computational tools for predicting protein folding kinetics.
Highlights
AlphaFold2 and RoseTTAfold could not directly identify unfoldable proteins.
The structural variations of computational models are correlated with protein foldability.
This correlation is independent of secondary structures.
The structural variations of computational models are quantitatively correlated with protein foldability but not protein function.
Competing Interest Statement
The authors have declared no competing interest.