Table 3:

Performance comparison of KaML-CBtree, KaML-GAT, PB, and alternative ML models^a

	PypKa		DeepKa		ANI-2X b		KaML-CBtree		KaML-GAT
	RMSE	CER	RMSE	CER	RMSE	CER	RMSE	CER	RMSE	CER
Asp	1.61 ± 0.28	56/917	1.23 ± 0.35	48/937	1.20 ± 0.14	52/929	0.75 ± 0.17	13/916	0.88 ± 0.17	5/901
Glu	0.86 ± 0.12	25/1039	0.84 ± 0.25	9/1068	0.81 ± 0.12	40/1075	0.60 ± 0.07	5/1076	0.76 ± 0.08	12/1053
His	1.13 ± 0.51	8/257	1.10 ± 0.49	12/248	0.52 ± 0.16	3/298	0.85 ± 0.14	11/209	0.86 ± 0.18	26/203
Cys	3.15 ± 0.97	21/56	n/a	n/a	n/a	n/a	1.50 ± 0.60	13/68	1.81 ± 0.56	16/59
Lys	1.01 ± 0.30	10/325	0.77 ± 0.25	2/322	1.14 ± 0.22	10/325	0.70 ± 0.21	1/325	0.87 ± 0.28	8/325
Tyr	1.49 ± 1.25	-	n/a	n/a	1.88 ± 1.46	-	1.24 ± 0.85	-	1.79 ± 0.95	-

^aPypKa¹² and ANI-2X³³ predictions were made with the local installed software provided by the authors. DeepKa predictions were obtained from the DeepKa web server.⁵⁷ n/a (not available) indicates that the model is unable to make predictions. CER of Tyr is not calculated due to the extremely small test sets (3 Tyr⁻).

^bOur test sets likely overlap with ANI-2X’s training set; removing overlap is impossible as the data in Ref³³ is unpublished.