Extended study on atomic featurization in graph neural networks for molecular property prediction

Table 14 Mean square errors on validation and test data of models selected with different scenarios

Dataset	Scenario	I	II	III	IV
ESOL (random)	arch. repr. val. \(\downarrow\) test \(\downarrow\)	2095 F 0.086 0.118	2095 \(\underline{\textsc {F}-\textsc {A}}\) 0.083 0.120	2095 F−A 0.083 0.120	2095 F−A 0.083 0.120
ESOL (scaffold)	arch. repr. val. \(\downarrow\) test \(\downarrow\)	3180 F 0.107 0.166	3180 F 0.107 0.166	\(\underline{1990}\) \(\underline{\textsc {F}-\textsc {A}}\) 0.106 0.221	1990 F−A 0.106 0.221
Rat	arch. repr. val. \(\downarrow\) test \(\downarrow\)	996 F 0.152 0.182	996 \(\underline{\textsc {A}\text {+}\textsc {H}}\) 0.147 0.196	996 A + H 0.147 0.196	996 A + H 0.147 0.196
Human	arch. repr. val. \(\downarrow\) test \(\downarrow\)	2095 F 0.143 0.218	2095 \(\underline{\textsc {F}-\textsc {C}}\) 0.141 0.213	2095 F−C 0.141 0.213	2095 F−C 0.141 0.213
QM9	arch. repr. val. \(\downarrow\) test \(\downarrow\)	914 F 4.531 9.193	914 \(\underline{\textsc {F}-\textsc {C}}\) 3.073 6.757	\(\underline{917}\) F−C 2.667 9.698	917 F−C 2.667 9.698

For each model we include information about which representation (repr.) and architecture (arch.) is selected. The representation search is performed over representations 1–12. Each change in architecture or representation when choosing a more expensive scenario is underlined

ISSN: 1758-2946