A Variational Autoencoder-Reinforcement Learning Framework for C13 NMR-Based Natural Product Structure Elucidation

Abstract

With almost two-thirds of all approved small-molecule drugs having natural products (NPs) and their derivatives as part of their development process, NPs undoubtedly play an important role in drug development. Nuclear magnetic resonance (NMR) spectroscopy is one of the most useful tools for analyzing the NP isolates and determining which are suitable for therapeutic applications. From a molecule’s NMR spectral data alone, chemists can often elucidate the molecule’s structure and determine its key properties. However, the NMR spectra of complex and large molecules, which are often the case for NPs, are typically hard to interpret because of overlapping signals and complex splitting patterns. Researchers have found ways to automate the process of elucidating structures from NMR spectral data. These methods include computing the structure through computational methods, predicting the structure from predicted substructures, and predicting the structure from the NMR spectral data directly. However, recent research has not attempted to elucidate structures that are as large and complex as NPs. This work focuses on the current progress of implementing and evaluating a machine learning framework that utilizes variational autoencoders (VAEs) and reinforcement learning (RL) to predict the structure of a potential NP from the NMR chemical shifts. A key advantage of this approach is addressing a limitation of existing generative models, which often struggle to generate complex molecules like NPs. Experienced researchers can often propose reasonable initial structural guesses based on NMR data, and this framework can automate the tedious optimization of their guesses to achieve the best match. Progress so far includes fine-tuning a VAE model for encoding a researcher's structural guess and decoding optimized structures, as well as developing an RL reward function that scores the quality of a guess based on predicted C13 NMR chemical shifts. Ongoing work involves integrating these components into a full RL framework and conducting comprehensive evaluations.