A Novel Molecular Representation Learning for Molecular Property Prediction with a Multiple SMILES-Based Augmentation

Table 2

The RMSE and MAE values of various approaches in ESOL, lipophilicity, and FreeSolv datasets. The predictive values of the approaches are partly derived from the related references [2, 8].

Model		ESOL		Lipophilicity		FreeSolv
		RMSE	MAE	RMSE	MAE	RMSE	MAE
CheMixNet	CNN_RNN	1.0419	0.8010	1.0513	0.8282	1.3553	1.0156
Conventional methods	XGBoost	0.9900	—	0.7990	—	1.7400	—
	Multitask	1.1200	—	0.8590	—	1.8700	—
Graph-based methods	MPNN	0.5800	—	0.7190	—	1.1500	—
	Weave	0.6100	—	0.7150	—	1.2200	—
3D-based models	Drug3D-Net	0.9683	0.7841	0.9930	0.8404	1.4709	1.1598
Our method (multiple SMILES)	RNN (one layer)	0.6585	0.5105	0.7929	0.6211	1.6051	1.2313
	RNN (two layers)	0.6394	0.4940	0.7960	0.6217	1.3575	1.0468
	CNN_RNN	0.5916	0.4448	0.7054	0.5481	1.0033	0.7859

The best results are highlighted in bold.