DP5 without DFT: uncertainty-calibrated graph neural net accelerates structure confirmation via NMR

Area of Science:

• Computational Chemistry
• Spectroscopy
• Machine Learning

Background:

• The DP4 and DP5 methods are established computational tools for assigning candidate structures to NMR spectra.
• These methods rely on Density Functional Theory (DFT) calculations, which are resource-intensive.
• Accurate structure assignment is crucial in various chemical research fields.

Purpose of the Study:

• To develop a faster alternative to existing DFT-based methods for NMR structure assignment.
• To introduce DP5q, a novel method utilizing machine learning to accelerate the process.
• To evaluate the accuracy and efficiency of DP5q compared to traditional methods.

Main Methods:

• Development of DP5q, employing a graph convolutional neural network and quantile regression.
• Replacement of computationally expensive DFT calculations with a machine learning model.
• Validation of DP5q on a large dataset of diverse molecular structures.

Main Results:

• DP5q significantly reduces calculation time compared to DFT-based DP5.
• The method achieves a modest decrease in accuracy while maintaining high performance.
• DP5q demonstrates efficacy on both large-scale and challenging structure assignment cases.

Conclusions:

• DP5q offers a computationally efficient and accurate solution for NMR spectral analysis and structure assignment.
• The integration of graph convolutional neural networks and quantile regression presents a promising advancement in computational chemistry.
• This rapid calculation method can broaden the applicability of structure assignment techniques in chemical research.