Real-time prediction of ¹H and ¹³C chemical shifts with DFT accuracy using a 3D graph neural network

Abstract

From quantum chemical and experimental NMR data, a 3D graph neural network, CASCADE, has been developed to predict carbon and proton chemical shifts. Stereoisomers and conformers of organic molecules can be correctly distinguished.

A framework for automated structure elucidation from routine NMR spectra

“…For each of the top n -ranked constitutional isomers identified by our model (with a user specified cutoff n ), the 1 H and 13 C peak shifts and coupling constants of all possible diastereomers could be predicted using quantum mechanical calculations 43–45 or recently developed machine learning protocols. 10,17,18,46 These predictions could then be compared with the experimental spectra to generate a ranked list of molecular structures with defined stereochemistry. A complementary approach is to expand our substructure prediction model to include substructures with defined stereochemical relationships using 3D fingerprints 47,48 ( e.g.…”

The DP5 Probability, Quantification and Visualisation of Structural Uncertainty in Single Molecules

“…DFT optimised geometries and NMR shift calculations for the molecules from NMRShiftDB were obtained from the training data of the GNN NMR shift prediction software CASCADE. 19 A single conformer of each of these molecules was optimised utilising the M062X functional and def2-TZVP basis set and NMR shift calculations performed using in 6-311g(d) basis set and mPW1PW91 functional. Calculation of FCHL atomic representations, l2 distances and gaussian kernel transformations were performed using the python package qml.…”

“…This dataset was originally developed for training machine learning models for NMR shift prediction, the generality and near chemical accuracy achieved by these models has been taken as justification for using this dataset in this similar task (details regarding this dataset can be found in the original publication). 19 It is well known that the expected magnitude and variance of DFT prediction errors for different functionals show strong complex, nonlinear dependencies on atomic environment. 52,53 This process takes this into account by weighting the contribution to the error PDF for the test atom of each atomic environment in the database by its similarity to the test environment.…”

“…The system was developed and rigorously tested utilising a dataset of 5140 organic molecules from NMRShiftDB originally selected for NMR prediction using machine learning by Paton et al (see supporting information section 6). 18,19 To demonstrate the performance of the DP5 probability in even more challenging situations, the system was also evaluated using 13 case studies of molecular structures that have undergone reassignments in the literature and addition 42 challenging relative stereochemistry elucidation examples. DP5 represents an exciting leap forward in quantifying molecular uncertainty.…”

The DP5 probability, quantification and visualisation of structural uncertainty in single molecules

“…The system was developed and rigorously tested utilising a dataset of 5140 organic molecules from NMRShiftDB originally selected for NMR prediction using machine learning by Paton et al (see supporting information section 6). 21,22 To demonstrate the performance of the DP5 probability in even more challenging situations, the system was also evaluated using 13 case studies of molecular structures that have undergone reassignments in the literature and addition 42 challenging relative stereochemistry elucidation examples.…”