Skip to main content

Advertisement

Fig. 2 | Journal of Cheminformatics

Fig. 2

From: Rapid prediction of NMR spectral properties with quantified uncertainty

Fig. 2

a A convolutional graph neural network computes per-vertex (atom) parameters by performing a weighted linear combination of neighboring vertices, and then passing the result through a nonlinearity. b Successive layers serve to aggregate information from more and more distant vertices (atoms), respecting the connectivity of the graph. The resulting per-vertex features are then passed through a series of linear layers to estimate a chemical shift value and a confidence level for each vertex (atom). c Mean absolute prediction error for \({^{13}\mathrm{C}}\) and \({^1\mathrm{H}}\) chemical shifts, comparing classical HOSE codes, ab initio calculations, and our graphical neural network. Error bars are bootstrap-estimated 95% confidence intervals for the mean. d GNN chemical shift errors for \({^{13}\mathrm{C}}\) as a function of true shift (ppm), showing mean error, standard deviation of the error, and max error. Lower panel shows the fraction of the training data present at that chemical shift. e Same as (d) but for \({^1\mathrm{H}}\)

Back to article page