Quantification of Molecular Structural Uncertainty Through Automated Computational Analysis of NMR Spectra

Abstract

Analytical data in the new era of automated synthesis are so extensive that rapid automatic processes are essential to gain information from the data. Manual methods of NMR analysis are incredibly labour intensive and prone to errors, whilst computational methods typically rely on DFT which is too computationally intensive for such high-through put use. In this work new ways of estimating the uncertainty in molecular structures through automated computational analysis of NMR data have been pioneered. This work is divided into five sections. The first section describes the development of DP4-AI, a program which automatically processes and analyses experimental NMR data to calculate the probability of each molecular structure in a list being the correct one given the correct structure is in the list. The second section discusses the characteristics of DFT-NMR prediction errors and shows how these prevent DP4-AI being applied to single molecular structures. The third section explains how new methods from the field of molecular machine learning can be employed to solve these problems. This discussion concludes in the fourth section with the development and evaluation of the DP5 Probability. The DP5 probability utilises principles from the field of molecular machine learning to calculate the probability of a single given molecular structure corresponding to an observed experimental NMR spectrum. The final section describes a complete reformulation of the DP5 probability and redesign of the DP5 program. This affords rapid and DFT-free calculation of the DP5 probability. Overall, the new methods developed and evaluated as part of this work bridge the gap between state-of-the-art robotic synthesis and structure determination and will help usher in the new era of automated chemical synthesis and discovery.