Revealing Nature's Blueprints for Single-Site Catalysis of C-H Activation with First-principles Modeling and Machine Learning

PROJECT ABSTRACT Enzymes catalyze selective small-molecule transformations with activity and selectivity that is seldom matched by non-biological catalysts. Metalloenzymes, such as non-heme iron enzymes (NHIEs), catalyze a diverse array of reactions involving C–H activation that are relevant to natural product biosynthesis and would be of great benefit if harnessed for medicinal chemistry. Understanding the role of the secondary sphere in catalysis is essential for describing how these enzymes work and for designing mimic catalysts capable of similar transformations under a wider range of conditions (i.e., pH and temperature) amenable to industrial catalysis. Computational modeling provides insight into the sources of enzymatic rate enhancement, the dynamics of substrates in the active site, and in the mechanism of biomimetic catalysts, all of which are challenging to determine experimentally. However, for difficult cases such as NHIEs, existing modeling techniques provide limited mechanistic insight because they are either too costly, too inaccurate, or require too much existing knowledge and user intervention. The overall vision for the PI's research program is to develop systematic methods and novel workflows to overcome cost–accuracy trade-offs in computational modeling for the discovery of new catalysts and understanding of enzymes. The PI has advanced the first machine learning (ML) models to discover new transition metal catalysts from millions of candidates, identifying opportunities to overcome trade- offs in catalyst performance. She has developed novel strategies for unveiling noncovalent interactions in NHIEs that determine their reaction selectivity and validated her predictions with experimental collaborators. The PI has advanced methods for making QM/MM systematic and robust and applied them to identifying contributions of rate enhancement in enzymes to determine where biomimetic counterparts fail. She has developed ML models to identify and avoid errors in first-principles modeling. The central hypothesis of the proposed research is that development of novel low-cost methods that enable the generation of larger datasets will reveal structure– property relationships in enzymatic and biomimetic C–H activation. The rationale is that dynamic effects and interactions with second-sphere residues that distinguish enzymes that catalyze different reactions cannot be understood without sufficient sampling and a broad comparison of behavior across the enzyme family. Over the next five years, the PI will 1) develop models for the prediction of regioselectivity in non-heme iron enzymes using neural network potentials to enable sampling, 2) systematically determine environmental contributions to catalysis, and 3) discover bioinspired homogeneous catalysts. The proposed research will produce a framework for predictive modeling in biological and bioinspired catalysis. The goals build upon methods the PI's lab has developed for modeling enzymes and non-biological systems and the recent results they have produced related to ML-accelerated catalyst screening and non-heme iron enzyme modeling.