Practical Machine Learning Strategies 4: Using Neural Networks to Replicate Proton and 13C NMR Chemical Shifts Obtained from ωB97X-D/6-31G* Density Functional Calculations

Abstract

Described are neural networks that accurately reproduce proton and ¹³C chemical shifts obtained from ωB97X-D/6-31G*//ωB97X-D/6-31G* density functional model GIAO calculations. They support uncharged, closed-shell molecules comprising H, C, N, O, F, S, Cl, and Br. Development involved training to ≈2.7 million equilibrium geometry and chemical shift calculations for a diverse collection of organic molecules (including synthetic drugs and natural products). Referenced to ωB97X-D/6-31G*//ωB97X-D/6-31G* calculations, chemical shifts from neural networks for 601 marine natural products show RMS errors of 0.05 ppm (proton) and 0.76 ppm (¹³C). RMS errors of 0.09 ppm (proton) and 1.02 ppm (¹³C) shifts result when equilibrium geometries from a previously described “estimated ωB97X-D/6-31G*” neural network model (trained to reproduce ωB97X-D/6-31G* geometries) are utilized. A second assessment of experimental ¹³C chemical shifts for 246 natural products is provided. Using neural network models to provide both geometries and chemical shifts: 45% of ¹³C shifts reproduce experimental values within 1 ppm, 73% within 2 ppm, and 86% within 3 ppm. Utilizing neural network models for both equilibrium geometries and chemical shifts reduces the computational time required for accurate proton and ¹³C chemical shifts from tens to hundreds of minutes to just a few seconds per molecule.