Prediction of the water solubility by a graph convolutional-based neural network on a highly curated dataset

Table 5 Performance of our GNN and other prediction tools on the two different test subsets (original data and new data, which were added)

Test set n = 980	Consensus GNN	EPI suite	OCHEM	ACD GALAS
Subset original dataset n = 756
Predictions possible for	756	712	756	756
r²	0.911	0.757	0.915	0.868
q²	0.908	0.620	0.913	0.863
rmse	0.630	1.303	0.612	0.770
bias	0.102	0.199	0.023	− 0.092
mne	− 2.58	− 7.15	− 3.25	− 5.74
mpe	3.23	5.80	3.31	4.53
95% neg	− 1.19	− 2.65	− 1.38	− 1.69
95% pos	1.45	3.14	1.31	1.50
Subset new dataset n = 224
Predictions possible for	224	222	224	224
r²	0.862	0.589	0.768	0.701
q²	0.845	0.254	0.749	0.685
*rmse*	0.744	1.627	0.947	1.060
bias	0.191	0.456	0.040	0.037
mne	− 3.32	− 5.12	− 4.99	− 5.58
mpe	2.18	5.38	3.74	3.54
95% neg	− 1.37	− 3.65	− 2.37	− 2.01
95% pos	1.48	3.69	1.87	2.19

The performance of the GNN compared to three other available prediction tools (EPI Suite, ACD GALAS and OCHEM) for the test set of 980 chemicals. The statistics for the test subsets of original data and novel data are given

ISSN: 1758-2946