A surrogate signal model for automated 1D 1H NMR compound quantification

Bioreactors are useful tools for bioprocessing and production of biologics, gene therapies and vaccines. Streaming data-driven process control systems can be valuable in lowering the cost of production or discovering novel reaction pathways. Nuclear Magnetic resonance (NMR) is an inexpensive spectroscopy technique that has characteristics that make it appropriate for on-line, high-throughput measurement of metabolic changes in a bioreactor vessel. Future quantitative NMR (qNMR) advancements for processing this type of streaming data could grant a unique possibility for in-situ bioprocessing applications. One significant challenge for 1D 1H qNMR is that the spectrum of a compound can deviate from its spectrum in a reference setting, especially across the various spectrometer frequency and concentration profile of metabolite mixture in the biofluid sample. A robust predictive or constraint model on the generative mechanism of the measured NMR signal can help guide future qNMR developments. We present an approximated 1D 1H NMR signal model that shows promise in fitting chemical shifts and other interpretable parameters for small mixtures of compounds. Our model use reference chemistry parameters of compounds to derive patterns between its nuclei via spin Hamiltonian simulations and hierarchical convex clustering on a spin angular momentum feature between the nuclei, which are quantum subsystems. These patterns are used to construct a surrogate model of the compound mixture with a lower degrees-of-freedom. Our approach does not require any phase or baseline correction techniques to pre-process the data, making it a generative model that fully accounts for the relative phase information, which is usually attenuated in a heuristic manner and ignored in conventional NMR data processing. We demonstrate the potential of this new methodology by fitting against real-world NMR reference compound experiments.