Fig. 2: Parity plots comparing reference values with predictions from the directed message passing neural network (D-MPNN) surrogate models on the test set.
From: Uncertainty quantification with graph neural networks for efficient molecular design
The color coding of the data points indicates the level of total uncertainty (\({\sigma }_{{{{\rm{total}}}}}^{2}\)) in the model predictions. Uncertainty quantification (UQ) across the panels varies: a–c ensemble and mean-variance estimation (MVE) methods were utilized; d–s the evidential method was applied in panels. The molecular structure similarity is calculated using the Tanimoto similarity metric. Abs. diff. of VEE absolute difference of vertical excitation energy, \({{{{\rm{R}}}}}^{2}\) (coefficient of determination), logP octanol-water partition coefficient, TPSA topological polar surface area. Source data is provided as a Source Data file.
