Table 1. Performance comparison of target prediction methods. The table gives the means and standard deviations of assay-AUC values for the compared algorithms and feature categories or input types. Overall, FNNs (second column) performed best. They significantly (α = 0.01) outperformed all other considered methods. The methods GC and Weave work directly on graph representations of compounds and SmilesLSTM uses the SMILES representations of compounds.

	FNN	SVM	RF	KNN	NB	SEA	GC	Weave	SmilesLSTM
StaticF	0.687 ± 0.131	0.668 ± 0.128	0.665 ± 0.125	0.624 ± 0.120
SemiF	0.743 ± 0.124	0.704 ± 0.128	0.701 ± 0.119	0.660 ± 0.119	0.630 ± 0.109
ECFP6	0.724 ± 0.125	0.715 ± 0.127	0.679 ± 0.128	0.669 ± 0.121	0.661 ± 0.119	0.593 ± 0.096
DFS8	0.707 ± 0.129	0.693 ± 0.128	0.689 ± 0.120	0.648 ± 0.120	0.637 ± 0.112
ECFP6 + ToxF	0.731 ± 0.126	0.722 ± 0.126	0.711 ± 0.131	0.675 ± 0.122	0.668 ± 0.118
Graph							0.692 ± 0.125	0.673 ± 0.127
SMILES									0.698 ± 0.130