Exploring data-driven strategies for modeling catalytic enantioselectivity and reactivity

The development of predictive understanding in chemical reactivity has historically relied on experimental observation and theoretical simulation. Recently, the integration with data-driven approaches has expanded the scope of chemical insight from qualitative case studies to quantitative, generalizable trends. This thesis investigates how data science and machine learning can be applied to reaction development, mechanistic interpretation, and optimization in organic chemistry. Chapter 2 examines the transferability of multivariate linear regression (MLR) models across catalytic systems. Models trained on chiral Brønsted acid and secondary amine catalysis are evaluated for their prediction ability in two data-limited environments. The first involves synergistic catalysis, where multiple catalysts participate in a single transformation. The second addresses reactions that generate complex products, such as those used in total synthesis. Using our statistical toolkit, I demonstrate the unexpected viability of transfer learning from data-rich systems to data-scarce regimes, despite subtle mechanistic differences or large structural deviations. Chapter 3 analyzes palladium-catalyzed decarboxylative allylic alkylation. The presence of mechanistic breaks in this system poses challenges for adapting the MLR. Nonlinear models are employed to capture these effects. Clusterwise linear regression, an algorithm developed by my colleague, is used to recover mechanistically interpretable relationships with the aid of density functional theory calculations, revealing an unexpected descriptor-dependent solvent effect. Chapter 4 addresses substrate matching and reactant generality through unsupervised learning analysis of reactant clusters across organocatalyzed transformations. Using k-means clustering of molecular descriptors and experimental yield data, I evaluate substrate performance across five major reaction classes, including Friedel-Crafts and Mannich-type transformations. This framework identifies substrates that confer broad reactivity across mechanistically distinct catalytic systems and establishes quantitative metrics for substrate generality, guiding substrate selection in reaction discovery and identifying privileged reactant scaffolds for optimization. Chapter 5 explores large language models (LLMs) for reaction prediction in organic chemistry. By fine-tuning transformer models on SMILES-based reaction representations and employing a chemistry-aware tokenizer, these models achieve high accuracy in forward and retrosynthetic prediction across diverse reactions. Moreover, I demonstrate that fine-tuned LLM embeddings can be used for quantitative analysis of functional group contributions to reactivity, providing mechanistic insights complementary to traditional physical organic approaches.