Table 2:

Performance metrics of KaML-CBtree for acid and base pK_a and protonation state predictions in comparison to the baseline models^a

	KaML-CBtree		KaML-GAT	PROPKA3		Null
	acid	base	acid + base	acid	base	acid	base
PCC	0.88 ± 0.03	0.92 ± 0.03	0.93 ± 0.02	0.74 ± 0.06	0.90 ± 0.04	0.55 ± 0.06	0.86 ± 0.04
RMSE	0.76 ± 0.13	0.79 ± 0.10	0.90 ± 0.08	1.28 ± 0.15	0.96 ± 0.18	1.36 ± 0.11	1.04 ± 0.10
MAXE	3.17 ± 0.61	2.60 ± 0.70	3.74 ± 0.49	3.72 ± 0.29	5.04 ± 0.46	5.55 ± 1.24	2.80 ± 0.67
Classification of protonation states at pH 7b
Pre (prot)	0.91	0.99	0.92	0.66	0.97	0.63	0.97
Rec (prot)	0.82	0.97	0.92	0.78	0.88	0.39	0.78
Pre (dep)	0.99	0.95	0.98	0.98	0.97	0.95	0.77
Rec (dep)	0.99	0.99	0.98	0.97	0.85	0.98	0.97
CERc	34/2099	12/536	70/2822	90/2055	53/618	141/2106	101/716

^aBaseline models include PROPKA3¹⁵ and null model which returns the model pK_a’s:⁵⁰ 3.7 for Asp, 4.2 for Glu, 6.5 for His, 8.5 for Cys, 9.5 for Tyr, and 10.4 for Lys.

^bPrediction is based on the probability of protonation given a predicted pK_a (see main text).

^cCritical error rate (CER) refers to the percentage of miss-classifying protonated as deprotonated or vice versa. All classification metrics were calculated after accumulating the predictions from all 20 holdout test sets.