Generative World Models to compute protein folding pathways

While there is an abundance of static data for the structure of biological macromolecules, the amount of data regarding their folding mechanisms and dynamics is scarce, posing a challenge to the training of AI models. World Models can come in handy in this regard. From limited data, they can build a latent, approximated representation of the spatiotemporal folding environment that can be used to train downstream AI models, overcoming data limitation issues. We developed a World Model-based generative framework to perform biomolecular simulations. The framework uses variational autoencoders to encode the protein structures and the dihedral moves into latent vectors. We train a feedforward neural network to predict the next latent structure after a latent move, allowing the model to develop its own understanding of the structural dynamics. Finally, in this “hallucinated” latent environment the framework trains an agent with evolutionary algorithms to learn a policy to drive folding simulations towards a target structure. Within the simulations, latent configurations can be decoded back into atomistic structures, allowing the model to be further regularized with Ramachandran-based constraints. Our framework can compute protein folding pathways four orders of magnitude faster (up to ~30000x) than standard MD. We have validated our results against the equilibrium MD data for fast-folding proteins. Finally, we illustrate the folding landscape of hCRBP2 in detail, identifying key folding intermediate states involved in its biology. Overall, the World Model facilitates the study of proteins by generative AI with applications in structure-based drug discovery.