(118f) You May Not Have Noticed, but Your Neural Network Did: Machine Learning from Simulated Enzyme Variants

Molecular simulation of enzymes has long been established as a valuable tool for making predictions about and developing rationalizations for observable behaviors such as reaction efficiency or thermostability. These (relatively) macroscopic observables are emergent properties of highly intricate interatomic relationships, but the exact nature of the relationships between particular subsets of the whole system and the observables of interest are most often too complex for humans to interpret. Nevertheless, it must be the case that given the appropriate framing and enough data, a suitable model could be trained to accurately quantify the relationship between each component of the system and the whole. Here, we train a machine learning model on molecular features from a large set of simulation data for variants of a particular engineered enzyme. We investigate the ability of the model to identify non-obvious relationships between components of the system, and to make testable predictions that help guide directed evolution experiments. Eventually, models that encode molecular-level understanding of enzyme structure-function relationships can be used to intelligently select desirable multiple mutants, even when the component single mutations are undesirable, thereby addressing a major shortcoming of experimental directed evolution.