AlphaFold is a novel method for 3D structure prediction based on deep learning. It uses a multiple sequence alignment of the modeled protein with its homologues for which amino acid sequences are known. These multiple sequence alignments intrinsically contain information about the 3D structure, in particular residue-residue distances. They are predicted and further converted into a structure model by AlphaFold. However, despite its success, AlphaFold has a limited capability to model the outcome of mutations, conformational changes, interactions with other types of molecules and other important phenomena. Here we used the output of AlphaFold in the form of inter-residue distance probability profiles to guide biomolecular simulations by metadynamics method. This approach was tested on folding of a mini-protein tryptophan cage. With parallel tempering metadynamics we were able to simulate multiple folding and unfolding events and to predict the temperature-dependent free energy profile in agreement with biophysical studies and reference simulations.
The project was supported by Czech Science Foundation (22-29667S).