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

Table 1 The reported GNN models in molecular property prediction

Year	Model Name	References	Datasets
			Regression (RMSE)			Classification (AUC_ROC)
			ESOL	FreeSolv	Lipop	MUV^a	HIV	BACE	BBBP	Tox21	ToxCast	SIDER	ClinTox
2019	Attentive FP^b	Xiong et al. [27]	0.503 ± 0.076	0.736 ± 0.037	0.578 ± 0.018	0.221 ± 0.047	0.832 ± 0.021	0.850 ± 0.012	0.920 ± 0.015	0.858 ± 0.014	0.805 ± 0.022	0.637 ± 0.017	0.940 ± 0.018
2019	D-MPNN^c	Yang et al. [24]	0.665 ± 0.052	1.167 ± 0.150	0.596 ± 0.050	0.122 ± 0.020	0.816 ± 0.023	0.878 ± 0.032	0.913 ± 0.026	0.845 ± 0.015	0.737 ± 0.013	0.646 ± 0.016	0.894 ± 0.027
2019	PAGTN^d	Chen et al. [29]	0.554 ± 0.060	NA	0.572 ± 0.040	NA	NA	0.880 ± 0.010	0.913 ± 0.030	NA	NA	NA	NA
2019	EIGNN^b	Chen et al. [28]	0.653 ± 0.025	1.273 ± 0.137	0.776 ± 0.071	NA	NA	NA	NA	NA	NA	NA	NA
2018	EAGCN^b	Shang et al. [30]	NA	0.950 ± 0.140	0.610 ± 0.020	NA	0.830 ± 0.010	NA	NA	0.860 ± 0.010	NA	NA	NA
2018	AGCN	Li et al [72].	NA	NA	NA	NA	NA	NA	NA	0.802	0.703	0.592	0.868
2017	GC^b	Wu et al. [32]	0.970 ± 0.010	1.400 ± 0.160	0.655 ± 0.036	0.046 ± 0.031	0.763 ± 0.016^e	0.783 ± 0.014^e	0.690 ± 0.009^e	0.829 ± 0.006	0.716 ± 0.014	0.638 ± 0.012	0.807 ± 0.047
2017	Weave^b	Wu et al. [17, 32]	0.610 ± 0.070	1.220 ± 0.280	0.715 ± 0.035	0.109 ± 0.028	0.703 ± 0.039^e	0.806 ± 0.002^e	0.671 ± 0.014^e	0.820 ± 0.010	0.742 ± 0.003	0.581 ± 0.027	0.832 ± 0.037
2017	DAG^b	Wu et al. [32]	0.820 ± 0.080	1.630 ± 0.180	0.835 ± 0.039	NA	NA	NA	NA	NA	NA	NA	NA
2017	MPNN^b	Wu et al. [32]	0.580 ± 0.030	1.150 ± 0.120	0.719 ± 0.031	NA	NA	NA	NA	NA	NA	NA	NA
2017	NA	Li et al [31]	NA	1.112	NA	NA	0.851	NA	NA	0.854	0.768	NA	NA

All the results were taken from the corresponding publication directly under the single model pattern and the best model for each dataset were italic
^aModel built on MUV was evaluated by AUC-PRC (the area under precision-recall curve); ^bAverage performance of 3 times independent runs with the standard deviation; ^cAverage performance of 10 times independent runs with the standard deviation except for HIV, and HIV is the average performance of 3 times independent runs with the standard deviation; ^dAverage performance of 10 times independent runs with the standard deviation; ^eModel was evaluated in scaffold splitting rather than random splitting; NA not available

ISSN: 1758-2946