Table 4.

Validation of the affinity prediction on different test sets, with WordPiece tokenizer for SMILES. Shown are Spearman ρ rank correlation coefficient, and mean-squared error (MSE) for the ensemble and the individual models. Values in parentheses indicate the uncertainty of the last reported digit.

	Ensemble		Model 1		Model 2		Model 3
Test set	ρ	MSE	ρ	MSE	ρ	MSE	ρ	MSE
Hold-out	0.864(3)	0.72(1)	0.830(3)	0.94(2)	0.846(3)	0.84(1)	0.854(3)	0.77(1)
Kinases (Davis et al., 2011)	0.30(1)	2.03(3)	0.23(1)	2.90(4)	0.26(1)	1.95(3)	0.30(1)	1.94(3)
Mpro (Achdout et al., 2020)	0.37(2)	1.23(5)	0.33(3)	1.23(5)	0.29(2)	1.49(5)	0.31(3)	1.44(5)
PLpro (Shen et al., 2021)	0.57(8)	0.71(10)	0.51(8)	0.45(7)	0.61(7)	0.76(10)	0.30(10)	1.7(2)