Explainable uncertainty quantifications for deep learning-based molecular property prediction

Table 3 Prediction accuracy and uncertainty performance of the atom-based uncertainty model (AtomUnc) and the molecule-based uncertainty model (MolUnc)

Dataset	Model	MAE	RMSE	ECE		ENCE
Dataset	Model	MAE	RMSE	Ale	Epi	Ale	Epi
QM9	AtomUnc	0.8904	2.7023	0.0129	0.0311	0.2120	0.4724
QM9	MolUnc	0.8394	2.7196	0.0700	0.0328	0.1772	0.5913
Zinc15	AtomUnc	0.00088	0.0083	0.0722	0.3284	0.1721	0.7238
Zinc15	MolUnc	0.00123	0.0113	0.3139	0.3201	0.6200	0.7051
Lipophilicity	AtomUnc	0.4191	0.5952	0.0396	0.2294	0.3704	1.5324
Lipophilicity	MolUnc	0.6709	0.8418	0.0119	0.3096	0.2441	1.9964
Delaney	AtomUnc	0.5031	0.6715	0.0622	0.0740	0.5578	0.6762
Delaney	MolUnc	0.5900	0.7520	0.1121	0.0203	0.5468	0.5314

ECE and ENCE are calculated based on aleatoric (Ale) and epistemic (Epi) uncertainties

ISSN: 1758-2946