Number and percentage of unique molecules obtained within different fingerprint-based similarity thresholds (δ) of the starting structures. The molecules in each experiment were generated from 250 000 random string mutations of the starting structures. Additionally, for celecoxib, we also formed the local chemical space with a scaffold constraint.

Starting structure (method)	Fingerprint	Number of molecules (and percentage)
		δ > 0.75	δ > 0.60	δ > 0.40
Aripirazole (SELFIES, random)	ECFP4	513 (0.25%)	4206 (2.15%)	34 416 (17.66%)
Albuterol (SELFIES, random)	FCFP4	587 (0.32%)	4156 (2.33%)	16 977 (9.35%)
Mestranol (SELFIES, random)	AP	478 (0.22%)	4079 (1.90%)	45 594 (21.66%)
Celecoxib (SELFIES, random)	ECFP4	198 (0.10%)	1925 (1.00%)	18 045 (9.44%)
Celecoxib (SELFIES, terminal 10%)	ECFP4	864 (2.02%)	9407 (21.99%)	34 187 (79.91%)
Celecoxib (SELFIES, central 10%)	ECFP4	111 (0.08%)	1767 (1.32%)	15 348 (11.45%)
Celecoxib (SELFIES, initial 10%)	ECFP4	368 (0.53%)	7345 (10.53%)	34 702 (49.74%)
Celecoxib (SMILES, random)	ECFP4	122 (18.43%)	515 (77.49%)	662 (100.00%)
Celecoxib (SMILES, terminal 10%)	ECFP4	90 (20.79%)	368 (84.99%)	433 (100.00%)
Celecoxib (SMILES, central 10%)	ECFP4	114 (22.18%)	419 (81.52%)	514 (100.00%)
Celecoxib (SMILES, initial 10%)	ECFP4	122 (19.71%)	490 (79.16%)	619 (100.00%)
Celecoxib (DeepSMILES, random)	ECFP4	132 (4.43%)	953 (31.99%)	2793 (93.76%)
Celecoxib (DeepSMILES, terminal 10%)	ECFP4	106 (9.73%)	513 (47.11%)	1083 (99.45%)
Celecoxib (DeepSMILES, central 10%)	ECFP4	53 (6.54%)	162 (19.98%)	658 (81.13%)
Celecoxib (DeepSMILES, initial 10%)	ECFP4	105 (9.28%)	609 (53.80%)	1106 (97.70%)
Celecoxib (SELFIES, scaffold constraint)	ECFP4	354 (0.44%)	6311 (7.79%)	53 479 (66.07%)
Celecoxib (CReM, ChEMBL: SCScore ≤ 2.5)	ECFP4	239 (0.58%)	5547 (13.47%)	14 887 (36.14%)