Fig. 13

Generalization performance on new conformations and unseen molecules. Results are shown for invDeepDFT (D=6, F=128), eqvDeepDFT (D=6, F=128), and LAGNet (D=4, F=128). Model accuracy is measured by NMAE(%). Two training splits were used: one with a single conformation per molecule and one with multiple conformations per molecule. Generalization is evaluated on three progressively harder test splits: the conformation split holds out new conformations of molecules seen during training; the structure split holds out entirely new molecules; and the scaffold split holds out molecules with unseen scaffolds