Fig. 3: Importance scores of substructures identified by MolGraph-xLSTM on the BBBP dataset.

For each molecule, the substructure with the highest model-assigned weight was analyzed using a random forest model to determine its relationship with BBBP labels. The substructure  − CC( = O)O − , containing a carboxylic group, received the highest importance score (highlighted by the blue dashed box).