Applications of machine learning and artificial intelligence in NMR
Stefan Kühn
Abstract
Nuclear magnetic resonance (NMR) spectroscopy was established in the 1940s, and Felix Bloch and Edward Mills Purcell were awarded a Nobel Prize for its development.1 Since then, the remarkable progress in instrumentation and applications has led NMR to permeate chemistry, biology and related fields. A challenge from the outset is that NMR measurements produced large amounts of data and have always strained both the memory and processing capacities of computers. This is true even for a one-dimensional spectrum of a pure compound but is even more true for multi-dimensional spectra, mixture analysis or reaction monitoring. Processing those data can be challenging. A prominent example is that of performing peak-picking to produce numerical peak lists, including a rich set of peak attributes, which can be easily used in subsequent analyses such as quantification, compound identification and structure determination. Using the full power of NMR data has been dictated in the past by the limits of human and computing machine capacity. Artificial intelligence (AI) and machine learning (ML) can advance the analysis of NMR data sets. With respect to terminology, we use AI here as a general term and include ML as probably the most prominent sub-set of it. Whilst definitions vary, ‘artificial intelligence’ is commonly described as computers being able to perform tasks requiring intelligence. Even though such descriptions are partly circular, defining intelligence by ‘requiring intelligence’, it is mostly intuitively clear what is meant. Since significant training and time from humans is needed to process NMR data, the idea to employ AI techniques in NMR spectroscopy is not new but has seen a renaissance in the past several years due to advancements in algorithms, availability of training data and increase computational power. This special issue presents the current state of applying ML and AI in various areas of NMR. A long-standing research area is shift prediction, where AI methods have been used and the application of convolutional neural networks (CNNs) recently gave a boost to the area. We cover this in a review (Jonas et al.). Since data are the basis of ML methods, the question of how to handle and store data is relevant. This area is represented by a demonstration of a simple, but machine-enabling, archiving system and its embedding in a world of FAIR (findable, accessible, interoperable and reusable) data in Rzepa et al. (this was accidentally published prematurely as Rzepa and Kuhn2). A lot of progress in the area was possible recently due to increased performance offered by computers and, in particular, by the use of Graphics Processing Units (GPUs) for training neural networks. Gill et al. give an overview of the role of these in chemistry and NMR from an industry perspective, focusing on recent successes and remaining challenges. Image analysis has made tremendous progress lately, mainly due to the application of CNNs. Kuhn et al. demonstrate that those techniques can be used to analyse spectral images directly for structural features of the compounds measured. Deng et al. identify structural features as well, but using the raw spectrum instead of spectra images, representing a different approach. Kim et al. use CNNs to identify metabolites in mixtures. Rozowski et al. address the problem of parameter estimation in biexponential models by using a neural network. Another important task in analytical chemistry is dereplication. Gerrard et al. demonstrate the use of the IMPRESSION prediction system for 15 N shift prediction. I would like to thank Professor Craig Butts, Professor David Rovnyak (special issues and features editors) and Paul Trevorrow (Wiley Executive Journals Editor) for giving me the opportunity to guest edit this special issue. Their support, enthusiasm and patience was fantastic. I am especially grateful to all of the researchers who kindly agreed to contribute, either by submitting articles or reviewing them, because without them, this special issue would not have been possible. The peer review history for this article is available at https://publons.com/publon/10.1002/mrc.5310.