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

Table 3 The performance comparison (RMSE) of the four descriptor-based and four graph-based models on the three regression 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
ESOL	1127	1	RMSE	SVM	0.158	0.624	0.516
				XGBoost	0.188	0.511	0.571
				RF	0.391	0.635	0.631
				DNN	0.448	0.568	0.553
				GCN	0.429	0.622	0.598
				GAT	0.402	0.518	0.604
				MPNN	0.467	0.546	0.665
				Attentive FP	0.407	0.479	0.471
FreeSolv	639	1	RMSE	SVM	0.347	0.423	0.674
				XGBoost	0.106	0.685	0.707
				RF	0.536	0.932	0.888
				DNN	0.483	0.527	0.724
				GCN	0.187	0.526	0.795
				GAT	0.496	0.634	0.851
				MPNN	0.316	0.772	1.050
				Attentive FP	0.529	0.517	0.813
Lipop	4200	1	RMSE	SVM	0.185	0.552	0.567
				XGBoost	0.145	0.524	0.556
				RF	0.481	0.625	0.649
				DNN	0.210	0.553	0.591
				GCN	0.315	0.573	0.612
				GAT	0.409	0.602	0.676
				MPNN	0.474	0.606	0.662
				Attentive FP	0.282	0.521	0.559