Could graph neural networks learn better molecular representation for drug discovery? A comparison study of descriptor-based and graph-based models

Table 4 The performance comparison (AUC_ROC) of the four descriptor-based and four graph-based models on the three classification datasets (data folds were generated from Attentive FP and the top three model were italic for each dataset)

Dataset	No.	Tasks	Metric	Model	Training	Validation	Test
HIV	40748	1	AUC_ROC	SVM	1.000	0.821	0.840
				XGBoost	0.999	0.842	0.848
				RF	0.962	0.805	0.846
				DNN	0.978	0.835	0.858
				GCN	0.994	0.862	0.857
				GAT	0.997	0.853	0.825
				MPNN	0.968	0.865	0.828
				Attentive FP	0.905	0.852	0.847
BACE	1513	1	AUC_ROC	SVM	0.976	0.883	0.861
				XGBoost	1.000	0.898	0.889
				RF	0.989	0.876	0.861
				DNN	0.973	0.921	0.883
				GCN	1.000	0.945	0.876
				GAT	0.996	0.937	0.848
				MPNN	0.972	0.921	0.848
				Attentive FP	1.000	0.923	0.889
BBBP	2035	1	AUC_ROC	SVM	0.988	0.922	0.899
				XGBoost	0.977	0.946	0.886
				RF	0.991	0.929	0.907
				DNN	0.981	0.933	0.856
				GCN	0.997	0.947	0.881
				GAT	0.999	0.947	0.872
				MPNN	0.944	0.961	0.889
				Attentive FP	0.971	0.952	0.907

ISSN: 1758-2946