Table 1 Comparison of methods’ performances on the datasets of S144, S129, and S99.

Dataset	Method	PCC	RMSE	AUC-ROC	AUC-PR	MCC
S144	PremPLI	0.46	0.77	0.78	0.36	0.40
	mCSM-lig¹⁷	0.41	0.91	0.75	0.31	0.33
	ML1¹⁹	0.12^a**	0.87	0.61*	0.20	0.20
	R15¹⁹	0.67**	0.72	0.77	0.50	0.52
	Prime²⁵	0.29	1.81	0.67	0.27	0.35
	FEP+²⁵	0.49	1.07	0.76	0.53	0.52
S129	PremPLI	0.48	0.78	0.81	0.35	0.39
S129	mCSM-lig¹⁷	0.20^b*	1.06	0.55*	0.28	0.40
S99	PremPLI^C	0.69	1.09	0.84	0.71	0.69
	A14²⁷	0.44**	1.35	0.78	0.55	0.54
	R14²⁷	0.33**	1.35	0.68	0.40	0.37
	RMD²⁷	0.48*	1.23	0.77	0.59	0.58

* and ** indicate statistically significant difference between PremPLI and other methods in terms of PCC (Hitter et al.⁵¹ test) and AUC-ROC (DeLong test) with p-value < 0.05 and p-value < 0.01, respectively. PremPLI^C: PremPLI was retrained after removing all mutations in the overlapped complexes with S99 from the training dataset. R15: Rosetta using the flex_ddg protocol and REF2015 scoring function. R14: Rosetta using the flex_ddg protocol and talaris2014 scoring function. A14: Amber14sb and GAFF(v2.1)/AM1-BCC force fields were used for proteins and ligands, respectively. RMD: the combination of R14 and A14. The results of more combinations are shown in Supplementary Fig. 7.
All correlation coefficients are statistically significantly different from zero (p-value < 0.01, t-test) except ^ap-value = 0.14 and ^bp-value = 0.02.

Quick links

Search