Investigations into Photosensitized Racemization of Atropisomers Guided by Density Functional Theory and Machine Learning

Abstract

While much work has been devoted to the enantioselective synthesis of atropisomers, methods for facilitating their enantiomerization remain underexplored. This thesis focuses on the use of computational chemistry and representation learning to elucidate the mechanism of photosensitized racemization of structurally diverse, sterically hindered biaryls and better understand their singlet and triplet properties. A survey of molecular representations and associated representation learning models is presented. Molecular representations are important for understanding and predicting properties of and relationships between chemicals. The advantages and disadvantages of common representations including strings, vectors, graphs, and coordinates are discussed. Simple linear and models are introduced in the context of molecular modeling along with nonlinear and representation learning techniques such as graph neural networks, recurrent neural networks, and transformers. The importance of selecting appropriate representations that capture relevant molecular information is emphasized, as it directly impacts the accuracy and interpretability of modeling efforts. The second chapter delves into the specific topic of photosensitized racemization of sterically hindered biaryls. Density functional theory (DFT) calculations are employed to study the singlet and triplet states of these biaryls, providing insights into their electronic structures and energy landscapes. Experimental results using iridium photosensitizers demonstrate the feasibility of photoracemization of clinically relevant heterobiaryls and lend legitimacy to computed rotational barriers. Systematic DFT investigations were conducted to examine steric and electronic trends. A simple parametric model was developed and found to accurately predict rotational barriers of biphenyls by evaluating steric contributions of each substituent. Additionally, calculations for a series of isosteric arenes with varying triplet energies revealed that rotational barriers decrease upon excitation in a manner proportional to the triplet energy and ground state barrier. A high throughput virtual screening approach was taken to develop a general model for predicting ground and excited state rotational barriers of biaryls. A representative dataset was generated using virtual cross coupling of commercially available fragments. A multi-tiered computational workflow was used to collect training data for a neural network model. Using a novel fingerprint representation that encodes localization information, transition state energies were estimated with high accuracy.