Prediction Of One Bond Carbon-Proton Coupling Constants (1JCH)


Development Of A Machine Leaning Method(Novel) To Optimise Quantum Chemical Calculations Using Density Functional Theory For The Prediction Of One Bond Carbon-Proton Coupling Constants (1JCH)


NMR spectroscopy is the most popular technique used for structure elucidation of small organic molecules in solution, but incorrect structures are regularly reported. One-bond protoncarbon J-couplings provide additional information about chemical structure because they are determined by features of molecular structure that are different from proton and carbon chemical shifts. However, these couplings are not routinely used to validate proposed structures because few software tools exist to predict them. We have assessed the accuracy of Density Functional Theory (DFT) predictions for 396 published experimental observations with the B3LYP functional and the TZVP basis set (1), using DFT calculations from the open-source software package NWChem. These results suggest that prediction of one-bond CH J-couplings (1JCH) using DFT is sufficiently accurate for structure validation. This will be of particular use in strained ring systems and heterocycles that have characteristic couplings and pose challenges for structure elucidation.


However, the quantum chemical calculations using NWChem needs to be optimised in terms of parameters, time-duration and computational resources to generate and analyse the data. A novel machine learning method need to be developed using pre-calculated 1JCH datasets using various molecular properties.


NWChem 1JCH predictions were performed on over 1410 diverse set of molecules (screened from eMolecules based on molecular diversity). Out of which only 927 Jobs were successful. A total of 9409 data points are available from the 927 molecules and these are mapped against the corresponding H atoms in the molecules and are stored as atomic properties in CML format.


Let us know if you are interest in analysing the data and develop a prediction model. We tried developing a Multi-Variate Linear Regression model, the results are promising. However, we believe that we can achieve better prediction accuracy if we can use the Novel Machine Learing Algorithms. A JAVA based web application is also in place to integrate any new prediction model developed.


(1) The Potential Utility of Predicted One Bond Carbon-Proton Coupling Constants in the Structure Elucidation of Small Organic Molecules by NMR Spectroscopy [DOI:10.1371 / journal.pone.0111576].

comments powered by Disqus