UC Berkeley

Catalysts lower overall reaction barriers without themselves being consumed; however, the precise properties enabling the desired activity and selectivity can be challenging to identify and exploit. In this dissertation we leverage synthesis coupled with data science-based toolkits to discover low-dimensional structure-activity relationships (SARs), linking simple catalyst properties to its performance. From the realized SARs, we can deploy machine learning algorithms to automate design, accelerating access to the most promising candidates.Chapter 1 – we provide a perspective on the physical organic basis of catalysis and molecular representations that seek to unite catalyst structure to its function. We reflect on strategies that unite simple representations with high chemical fidelity. Together, we expect these goals to improve predictive and rational design under complex settings. Chapter 2 – we present a data-driven approach to interrogate the activity and selectivity of a supramolecular cage-mediated cascade reduction. Data science-based inquiry identifies univariate regressions and classifications that precisely quantify the emergence of size-exclusion across the higher-order assembly within the host cavity. Moreover, simple and experimentally validated surrogates for competing, attractive non-covalent interactions are identified. Chapter 3 – we deploy statistical inquiry to identify a single descriptor – active site-based buried volume – is uniquely relevant for the construction of models for stereoselectivity by diverse sets Brønsted acid organocatalysts. We show the representation transcends steric inference and likely accommodates a diverse array of stereoelectronic interactions responsible for stereoinduction. Chapter 4 – we generalize the relationship between active site-based buried volume and stereoselectivity by Brønsted acid organocatalysts. We present a holistic interpretation based on complementarity where a chiral catalyst preferentially matches one chiral transition state. From this perspective, we construct predictive Gaussian regression models that generalize stereoselectivity beyond empirical examples onto new, disparate and unfamiliar solutions. Chapter 5 – we conclude our extensive investigation on generalizing stereoselectivity of Brønsted acid organocatalysts by presenting an algorithm for the design of stereoselective solutions. The leverages active learning-based machine learning to efficiently navigate chemical space, affording the inverse design of highly stereoselective Brønsted acid organocatalysts. Chapter 6 – we expand the reaction manifold of Brønsted acid organocatalysts to a suite of open-shell reactions, including: group transfer, alkene additions, and excited state quenching paradigms. Dithiophosphoric acids are uniquely capable of accessing these reaction modes and we conduct rigorous mechanistic inquiry to rationalize this. From these foundations, we expect our findings will enable data-driven investigations into new, stereoselective open-shell reactions.