"Development of language modeling for chemical analysis and design"
Generative modeling techniques offer the opportunity to achieve intelligent and efficient molecular analysis and design, making their development critically important in areas ranging from small-molecule lead optimization to materials discovery. Language modeling techniques in particular have proven effective at navigating highly combinatorial, discrete design spaces, suggesting that similar approaches could transform molecular design. In this dissertation, I develop language-model-based methods for molecular applications. Beginning with the problem of property-directed generative design, I present an energy-based alignment algorithm for chemical language models that yields highly scalable and sample-efficient molecular generators. Next, I apply chemical language models to spectroscopic analysis by designing, implementing, scaling, and post-training an end-to-end deep learning model for small-molecule structure elucidation from one-dimensional NMR spectra, achieving accurate and efficient elucidation for systems with as many as 40 non-hydrogen atoms. I then extend this framework with retrosynthetic context to enable context-aware structure elucidation, substantially improving accuracy without degrading performance when only spectra are available. Finally, I demonstrate how multifidelity asynchronous reinforcement learning with large language models (LLMs) can achieve frontier-level performance in small-molecule lead optimization against an expensive in silico oracle. Together, these projects explore how neural sequence models can inform both the understanding and the practice of chemistry.