Predicted active and inactive docked poses for testing set active compounds based on GA/SHAP-selected features.

Compoundsa	Predicted number of active and inactive docked poses
	Xgboost			RF
	Active poses	Inactive poses	% Active posesb	Active poses	Inactive poses	% Active posesb
10	105	14	88.2	98	21	82.4
22	113	21	84.3	99	35	73.9
34	86	22	79.6	88	20	81.5
49	76	20	79.2	73	23	76.0
54	60	29	67.4	57	32	64.0
66	76	20	79.2	75	21	78.1
72	98	23	81.0	87	34	71.9
78	77	2	97.5	65	14	82.3
107	54	17	76.1	58	13	81.7
113	57	6	90.5	56	7	88.9
115	90	2	97.9	73	19	79.3
116	67	22	75.3	65	24	73.0
126	97	2	98.0	80	19	80.8
127	66	1	98.5	55	12	82.1
133	51	2	96.2	41	12	77.4
135	65	0	100	56	9	86.2
146	82	5	94.3	73	14	83.9
148	54	3	94.7	45	12	78.9
156	90	7	92.8	85	12	87.6
162	62	5	92.5	59	8	88.1
250c	10	32	23.8	24	18	57.1
312	0	1	0	1	0	100

^aCompounds' numbers and bioactivities are as in Table S1.

^bDetermined by dividing the number of poses labeled as “active” by the total number of poses (labeled as “active” and “inactive”).

^cThe percent active poses of this compound (IC₅₀ = 5000 nM) was used as threshold to classify screened compounds into potential active and inactive STAT3 inhibitors in both GA-RF and GA-XGboost models.