Pretraining graph transformers with atom-in-a-molecule quantum properties for improved ADMET modeling

Table 2 Results of the fine-tuning on internal microsomal clearance dataset

	Metric	Scratch	All	Charges	Nmr	Fukui_n	Fukui_e	Masking	Homo-lumo
Clearance_1	\(R^2\) \(\uparrow\)	0.505 ± 0.010	0.640 ± 0.004	0.629 ± 0.006	0.635 ± 0.006	0.599 ± 0.004	0.593 ± 0.004	0.580 ± 0.012	0.602 ± 0.006
	Spearman \(\uparrow\)	0.728 ± 0.008	0.807 ± 0.003	0.799 ± 0.004	0.801 ± 0.003	0.785 ± 0.003	0.785 ± 0.001	0.774 ± 0.007	0.786 ± 0.004
Clearance_2	\(R^2\) \(\uparrow\)	0.534 ± 0.006	0.653 ± 0.004	0.633 ± 0.003	0.643 ± 0.005	0.598 ± 0.007	0.610 ± 0.008	0.597 ± 0.002	0.607 ± 0.005
	Spearman \(\uparrow\)	0.750 ± 0.005	0.818 ± 0.003	0.807 ± 0.004	0.811 ± 0.002	0.789 ± 0.002	0.795 ± 0.006	0.786 ± 0.002	0.794 ± 0.002

Results are reported for both values of clearance in the dataset and for all pretraining strategies both in terms of \(R^2\) coefficient and in terms of Spearman’s rank coefficient. Highlighted values denote the best performance achieved among our models, based on the average value and t-tests paired across seeds

ISSN: 1758-2946