Table 3. Overall Performance for Classification Tasks on Toxicity-Related Datasets of MoleculeNet^3a.

Data set ↑	Tox21	ToxCast	SIDER	ClinTox	BACE	BBBP
No. molecules	7,677	8,405	1,356	1,438	1,511	1,959
Label dist.	0.06:0.77:0.17	0.03:0.27:0.70	0.57:0.43:-	0.51:0.49:-	0.54:0.46:-	0.76:0.24:-
No. tasks	12	617	27	2	1	1
D-MPNN	0.759 ± 0.007	0.655 ± 0.003	0.57 ± 0.007	0.906 ± 0.006	0.809 ± 0.006	0.724 ± 0.004
AttentiveFP	0.761 ± 0.005	0.637 ± 0.002	0.606 ± 0.032	0.847 ± 0.003	0.784 ± 0.022	0.643 ± 0.018
GEM	0.781 ± 0.001	0.692 ± 0.004	0.672 ± 0.004	0.901 ± 0.013	0.856 ± 0.011	0.724 ± 0.004
SMILES-T	0.691 ± 0.011	0.578 ± 0.011	0.504 ± 0.028	0.819 ± 0.045	0.739 ± 0.075	0.931 ± 0.012
ET (single)	0.780 ± 0.004	0.685 ± 0.009	0.606 ± 0.01	0.851 ± 0.027	0.832 ± 0.009	0.960 ± 0.03
ET (multi)	0.789 ± 0.003	0.623 ± 0.008	0.560 ± 0.011	0.843 ± 0.012	0.816 ± 0.013	0.955 ± 0.008

^aThe normalized label distribution is denoted as active:inactive:nan. The equivariant transformer is denoted as ET. Here, (single) and (multi) differentiate between single- and multi-conformer training. We report the standard deviations for five different seed runs as subscripts. Two numbers are written in bold in one column if the standard deviations overlap.