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

Abstract

The evaluation and assignment of candidate structures to NMR spectra can be facilitated by the DP4 method, which assumes that one of the candidate structures is correct, and the DP5 method, which calculates the probability of a correct assignment for each candidate individually. Both of these methods require DFT calculations and so a significant amount of computer resource. In this paper we present DP5q, a new version of DP5 which uses a Graph Convolutional Neural Network and quantile regression to replace the DFT-based algorithm. This dramatically increases the speed of the calculation at the cost of a modest decrease in accuracy. We demonstrate the efficacy of this rapid calculation both on a test set of thousands of molecules and also on cases selected for the difficulty of assigning structure.