Deep learning-based approaches have advanced the de novo design of static proteins and enabled robust prediction of structure from sequence. However, computational approaches for designing specific protein structures with user-defined features important for function, such as complex dynamics and cooperative transitions, are still limited. Guo et al. have now reported a method to design proteins capable of dynamic motion triggered by binding. Their approach uses deep learning to guide the search of sequence and structure space along with molecular dynamics simulations to provide a greater understanding of protein motions.

The team started with a single-state Ca2+-binding protein (derived from the N-terminal domain of troponin C) and aimed to design a protein that could adopt multiple conformational states. They first used Rosetta to design a sequence predicted to adopt an alternative structure in the targeted part of the protein. Next, they used computational mutational scanning to define the minimal set of residues required to define the two states. This enabled them to increase the sequence identity in sites that, when mutated, did not perturb the structures, and generate designs that were similar but predicted to have different populations of states. These designs were experimentally validated by NMR and molecular dynamics. The team found that mutation at a single position was enough to tune the distribution of populated states.