Reconstruction of lossless molecular representations from fingerprints

Table 1 Translation-related statistics regarding the domain-specific datasets generated by the structural fingerprints used for the performance analysis, together with the targeted molecular representations, SMILES, and SELFIES

Abbreviations	Description	Dim	Sequence length		Token size
Abbreviations	Description	Dim	Ave.	Max	Token size
Predefined substructures
MACCS		166	50	107	160
Paths and feature classes
Avalon	Hashed	512	182	470	516
Path-based
HashAP	Atom pair - hashed	2048	92	273	1998
RDK4	RDkit fingerprint - hashed	2048	83	288	2052
RDK4-L	RDK4 - with no branch	2048	58	209	2052
4-atom-paths
TT	Topological torsion	sparse	32	124	54973
HashTT	TT - hashed	2048	31	118	2052
Circular
AEs	Morgan radius 1	sparse	29	65	54076
ECFP0	Morgan radius 0 - hashed	2048	10	25	100
ECFP2	Morgan radius 1 - hashed	2048	28	64	2052
ECFP4	Morgan radius 2 - hashed	2048	47	103	2052
FCFP2	Feature-class of ECFP2	2048	20	51	1576
FCFP4	Feature-class of ECFP4	2048	36	86	2052
Unique Representation
SMILES	Tokenized atom-wise		51	125	109
SELFIES	Generic tokenization		44	127	205

ISSN: 1758-2946