Table 3:

Full regression and classification results of three characterization tasks on the four enzyme datasets. Baseline model here used Morgan bit vector encoding and two layers of fully connected MLPs.

Tasks	Dataset	Phosphatase	Halogenase	Kinasea	Aminotransferase
	#Seq. × #Subs.	218 × 168	42 × 62	318 × 72	25 × 18
Simple task	R Conv+ECFP6	0.816	0.892	0.845	0.838
	R Baseline	0.728	0.838	0.805	0.808
	AU-PRC	0.710	0.732	0.809	0.867
	ROC-AUC	0.901	0.937	0.905	0.905
Substrates task	R Conv+ECFP6	0.681	0.545	0.335	0.470
	R Baseline	0.649	0.521	0.205	0.322
	AU-PRC	0.588	0.606	0.403	0.756
	ROC-AUC	0.858	0.931	0.730	0.697
Sequence task	R Conv+ECFP6	0.465	0.673	0.735	0.790
	R Baseline	0.422	0.581	0.716	0.796
	AU-PRC	0.418	0.743	0.745	0.790
	ROC-AUC	0.695	0.909	0.889	0.842

^aClassification performed to Kinase dataset uses self-defined labels.