Table 3.

Distributional results on QM9. CharacterVAE⁴⁹, GrammarVAE⁵⁰, GraphVAE²³ and MolGAN⁵¹ results are taken from Cao and Kipf⁵¹.

	Model	Valid	Uniq	Novel	KL Div	Fréchet Dist
SMILES	CharacterVAE	0.103	0.675	0.900	N/A	N/A
	GrammarVAE	0.602	0.093	0.809	N/A	N/A
	LSTM (ours)	0.980	0.962	0.138	0.998	0.984
	Transformer Sml (ours)	0.947	0.963	0.203	0.987	0.927
	Transformer Reg (ours)	0.965	0.957	0.183	0.994	0.958
Graph	GraphVAE	0.557	0.760	0.616	N/A	N/A
	MolGAN	0.981	0.104	0.942	N/A	N/A
	NAT GraphVAE (ours)	0.875	0.317	0.895	0.843	0.509
	MGM (ours proposed)	0.886	0.978	0.518	0.966	0.842

NAT GraphVAE²⁵ stands for non-autoregressive graph VAE. Models labelled as ‘ours’ were trained by us and subsequently used to carry out generation. Our masked graph model results correspond to a 10% masking rate and training graph initialization, which has the highest geometric mean for all five benchmark metrics. (See the Supplementary Discussion section of the Supplementary Information for details.) Values of validity(↑), uniqueness(↑), novelty(↑), KL Div(↑) and Fréchet Dist(↑) metrics are between 0 and 1.