Evaluating uncertainty-based active learning for accelerating the generalization of molecular property prediction

Table 1 Solubility uncertainty estimation evaluation

UQ type	Method	Model	ENCE	\(\rho _{\text {error}}\)	\(\rho _{\text {ood}}\)	\(\rho _{\Delta \text {error}}\)
Baseline	GBM	GBM	0.098	\(\mathbf {0.293 \pm 0.002}\)	−0.191	0.386
Ensemble	MCDO	MDM	1.585	0.180	−0.219	0.404
	MCDO	GNN	1.951	0.109	−0.091	0.111
	Ensemble	MDM	2.349	0.296	−0.099	−0.142
	Ensemble	GNN	2.830	−0.010	−0.145	0.264
Target value	Evidential	MDM	1.147	0.381	0.103	−0.404
	Evidential	GNN	0.457	0.145	0.207	0.142
	MVE	MDM	\(\mathbf {0.278 \pm 0.034}\)	0.378	−0.037	−0.261
	MVE	GNN	\(\mathbf {0.112 \pm 0.026}\)	0.041	0.035	−0.314
Union	GBM	MDM	0.366	0.111	−0.371	−0.618
Union	GBM	GNN	0.126	0.278	−0.352	0.179
Distance	Data density (FP)	MDM	–	0.151	0.965	0.857
		GNN	–	0.143	0.965	0.693
		GBM	–	0.124	0.965	0.546
	Data density (EB)	MDM	–	0.183	0.500	0.818
		GNN	–	0.178	0.500	0.679
		GBM	–	0.142	0.500	0.539
Consensus	GBM, MCDO, MVE	MDM	0.230	0.313	–	–

The uncertainty approach with best mean performance across models for each metric is shown in bold. For the density method, FP refers to fingerprint-based similarity and EB refers to embedding-based similarity

ISSN: 1758-2946