AbstractThe primary challenge of fixed-backbone protein design is to find a distribution of sequences that fold to the backbone of interest. This task is central to nearly all protein engineering problems, as achieving a particular backbone conformation is often a prerequisite for hosting specific functions. In this study, we investigate the capability of a deep neural network to learn the requisite patterns needed to design sequences. The trained model serves as a potential function defined over the space of amino acid identities and rotamer states, conditioned on the local chemical environment at each residue. While most deep learning based methods for sequence design only produce amino acid sequences, our method generates full-atom structural models, which can be evaluated using established sequence quality metrics. Under these metrics we are able to produce realistic and variable designs with quality comparable to the state-of-the-art. Additionally, we experimentally test designs for a de novo TIM-barrel structure and find designs that fold, demonstrating the algorithm’s generalizability to novel structures. Overall, our results demonstrate that a deep learning model can match state-of-the-art energy functions for guiding protein design.SignificanceProtein design tasks typically depend on carefully modeled and parameterized heuristic energy functions. In this study, we propose a novel machine learning method for fixed-backbone protein sequence design, using a learned neural network potential to not only design the sequence of amino acids but also select their side-chain configurations, or rotamers. Factoring through a structural representation of the protein, the network generates designs on par with the state-of-the-art, despite having been entirely learned from data. These results indicate an exciting future for protein design driven by machine learning.
A Combinatorial Toolbox for Protein Sequence Design and Landscape Analysis in the Grand Canonical Model