Table 3.

Table for comparing metrics for the three metrics at a temperature of 0.8 for 10k generated molecules. All metrics were evaluated with these 10k molecules, except uniqueness at 1k

Model	Condition type	Interval	Novelty [%] $\uparrow$	Uniqueness @ 1k [%] $\uparrow$	Uniqueness [%]$\uparrow$	Validity [%] $\uparrow$	MAD $\downarrow$
Llamol	Unconditional		89.7	100.0	99.9	99.5
	LogP	[-2, 7; 1]	86.5	100.0	99.9	99.3	0.159
	LogP	{2, 4, 6}	85.5	100.0	99.7	99.42	0.191
	SAScore	[1, 10; 1]	85.4	100.0	98.8	82.1	0.390
	SAScore	{2, 3, 4}	86.6	100.0	99.9	99.7	0.103
	Molecular weight	[1, 10; 1]	88.8	100.0	99.3	97.5	0.157
	Molecular weight	{2, 3, 4}	84.3	100.0	99.6	99.45	0.041
MolGPT	Unconditional (MOSES)		79.7		100.0	99.4
	Unconditional (GuacaMol [51])		100.0		99.8	98.1
	LogP$_{}^1$	{2, 4, 6}	100.0		99.8	97.1	0.23
	SAScore$_{}^1$	{2, 3, 4}	100.0		99.5	97.7	0.13
MolGPT (relative attention) [53]	Unconditional (MOSES)		87.9		100.0	99.2
	Unconditional (GuacaMol		100.0		100.0	97.8
	LogP$_{}^1$	{2, 4, 6}	100.0		100.0	96.9	N/A
	SAScore$_{}^1$	{2, 3, 4}	100.0		99.7	98.6	N/A
Transformer-Decoder Generator [54]	Unconditional (MOSES 30k$_{}^2$)		97.38		99.92	91.15
GraphGPT [49]	Unconditional (MOSES)		78.7		99.9	99.5
	Unconditional (GuacaMol)		100.0		99.9	97.5
	LogP$_{}^1$	{2, 4, 6}	100.0		99.8	96.9	0.22
	SAScore$_{}^1$	{2, 3, 4}	100.0		99.6	97.7	0.14

$_{}^1$ Trained on the GuacaMol dataset

$_{}^2$ The experiments of the Transformer-Decoder Generator (FSM-DDTR) were conducted on the MOSES dataset with 30k generated molecules instead of 10k