Fig. 3: Machine learning approach and performance.
From: DyeDactic workflow to predict halochromism of biosynthetic colourants
A MPNN approach to predict λmax for colourants taking into account solvent and colourant structure. B A parity plot between calculated lowest vertical transition energies for the natural product data set. Systematic correction is not applied and the red line is the bisector illustrating perfect correlation. ‘Perceived colour of light’ palette provides a correspondence between the light of the given energy and its colour. C Neural network loss changes during training. Training and test set losses are shown in blue and red, respectively. Green dashed line marks learning rate changes over training (axis on the right belongs to it). Training and validation error uncertainties were estimated based on five-fold cross-validation (transparent blue and red). D Prediction results for the validation set composed of artificial dyes (yellow to purple; yellow marks the highest density for points) and E prediction results for the natural colourants test set, the adopted colour scheme is the same as in (D). The colour bar on the right shows the perceived colour of light with energy on the y-axis. Structures of the two outliers are shown.
