Fig. 6

Relative importance given to expert-derived structural alerts. Relative token importance given to atoms corresponding to expert-derived structural alerts. Relative importance is given for canonical representations of different XAI methods of different training settings. Training settings include untrained, frozen encoder (native), completely random model (untrained), distances of the training data (train), distances of the test set on a random split model (random), enumerated training (enumerated), the statistics of the CNN model (CNN) and the averaged values of all SMILES enumerations of the ME2C model