Table 2. Overall Performance for Classification and Regression Tasks on Toxicity-Related Data Sets of TDCommons³⁸ and ToxBenchmark^37a.

Data set	Ames ↑	hERG ↑	DILI ↑	Skin Reaction ↑	LD50 ↓	Tox- Benchmark ↑
No. molecules	7,269	650	470	403	7,353	6,489
Label dist.	0.55:0.45	0.68:0.32	0.5:0.5	0.68:0.32	2.54/0.95	0.53:0.47
AttrMasking	0.842 ± 0.008	0.778 ± 0.046	0.919 ± 0.008	-	0.685 ± 0.025	-
AttentiveFP	0.814 ± 0.008	0.825 ± 0.007	0.886 ± 0.015	-	0.678 ± 0.012	-
Fingerprint-based	0.865 ± 0.002	0.875 ± 0.003	0.937 ± 0.004	-	0.588 ± 0.005	0.86 ± 0.01
SMILES-T	0.697 ± 0.011	0.703 ± 0.056	0.760 ± 0.041	0.633 ± 0.051	0.715 ± 0.012	0.720 ± 0.014
ET (single)	0.836 ± 0.003	0.839 ± 0.017	0.878 ± 0.013	0.662 ± 0.033	0.653 ± 0.008	0.881 ± 0.008
ET (multi)	0.804 ± 0.004	0.763 ± 0.021	0.885 ± 0.030	0.581 ± 0.055	0.660 ± 0.01	0.879 ± 0.005

^aFor Ames, hERG, DILI, Skin Reaction, and ToxBenchmark, we report the normalized label distribution as active:inactive. For LD50, the mean and standard deviation is given. The equivariant transformer is denoted as ET. Here, (single) and (multi) differentiate between single- and multi-conformer training. We report the standard deviation for five different seed runs in subscripts. Two numbers are written in bold in one column if standard deviations overlap.