Fig. 4

Analysis of predictive performance and map structure. a The performance of the direct semantic DirSem and the imputed semantic ImpSem models (open blue circles and squares, respectively) as the number of descriptors used during training is increased. b Prediction performance for each molecule, as measured by average correlation across descriptors between the ground truth ratings and the ratings predicted by the DirSem and mixed models (blue and green dots, respectively). The best-predicted molecules are toward the bottom of the chart, limit of significance (for DirSem model) is shown by the dotted gray line. c Histograms showing the median correlation across molecules for each descriptor for left the DirSem model and right the mixed model. d Odor wheel: prediction performance for each descriptor—as measured by the correlation across molecules between the ground truth and the predictions from the DirSem model—is indicated by the color of the text (see bar for scale). Descriptors are arranged and clustered based on their semantic vectors