Table 2.

Comparison of generator quality metrics and various other parameters of proposed model with other baseline architectures.

Architecture Type	Model Name	Dataset used	Size of Molecule	No.of Trained Molecules	No.of Generated Molecules	Generator quality metrics	Task
RNN and AE based Architecure	Grammar VAE [50]	ZINC	<39 heavy atoms	250,000	100,000	7.2% (V)	Penalized logP
	SD VAE [51]	ZINC	<39 heavy atoms	250,000	100,000	43.5% (V)	Penalized logP
	AAE [52]	ChEMBL	<121 characters	1.3 million	no data	77.4% (V)	Drug analog generation
	ECAAE [53]	ZINC	<58 characters	1.8 million	10000	No data	Structural analogs
GAN and RNN based Architecure	ORGAN [54]	QM9	<52 characters	5000	No data	80.3% (V)	nlogP
	ORGANIC [55]	QM9	<10 heavy atoms	5000	No data	0.2–99%(V), 86% (N)	nQED
	ATNC [56]	ChemDiv	<91 characters	15000	157986	72%(V),77% (N)	No.of unique heterocycles
	RANC [57]	ChemDiv	<91 characters	15000	896000	58%(V),48% (N)	No.of unique heterocycles
RNN based Architecure with RL	REINVENT [58]	ChEMBL	10 - 50 heavy atoms	1.5million	12800	94% (V), 90% (N)	Drug analog generation
	ReLeaSE [59]	ChEMBL	No data	1.5 million	1 million	95%(V),95.3% (N)	Inhibitor of JAK2
	ChemTS [60]	ZINC	No data	250,000	No data	No data	Penalized logP
RNN based Architecure	Segler et al. [33]	ChEMBL	No data	1.4 million	976,327	97.7% (V), 89.4% (N)	Plasmodium falciparum,5 - HT2A
	Bjerrum et al. [61]	ZINC	No data	1,611,889	50000	98% (V), 63% (N)	Retro-synthetic route of easy/medium/hard group
	Gupta et al. [16]	ChEMBL	34-74 characters	541,555	30107	93% (V), 92% (N)	PPARs, Trypsin
	Ours	ChEMBL and MOSES	34-128 characters	2.9 million	10000	70.50% (V), 99.83% (N),98.99% (U)	Inhibitor of 3CLPro (novel Corona virus main protease)

Q_valid, Q_Unique, Q_Novel are represented as V, U, N respectively. Autoencoders (AE), Adversarial autoencoder (AAE), Generative Adversarial network (GAN).