Table 5.

Conditional generation results on QM9. Results for MGM are chosen from a range of sampling iterations and both initialization strategies.

Target Condition	Model	G-mean	Unique Count	Property Value	KLD Score
MolWt = 120	NAT GraphVAE	0.623	3048	124.47 ± 7.58	0.843
	MGM	0.522	8800	120.02 ± 7.66	0.811
	MGM - Final Step	0.404	8509	119.42 ± 7.67	0.761
	Dataset	—	—	—	0.679
MolWt = 125	NAT GraphVAE	0.565	2326	127.21 ± 7.05	0.827
	MGM	0.561	9983	125.00 ± 8.48	0.850
	MGM - Final Step	0.354	9293	122.48 ± 7.20	0.936
	Dataset	—	—	—	0.835
MolWt = 130	NAT GraphVAE	0.454	1204	129.12 ± 6.79	0.614
	MGM	0.501	9465	128.85 ± 8.85	0.705
	MGM - Final Step	0.369	8892	126.85 ± 7.43	0.789
	Dataset	—	—	—	0.695
LogP = -0.4	NAT GraphVAE	0.601	2551	−0.409 ± 0.775	0.739
	MGM	0.424	9506	−0.349 ± 0.503	0.803
	MGM - Final Step	0.300	9495	−0.337 ± 0.523	0.876
	Dataset	—	—	—	0.811
LogP = 0.2	NAT GraphVAE	0.562	2188	0.051 ± 0.746	0.803
	MGM	0.378	9524	0.200 ± 0.468	0.846
	MGM - Final Step	0.376	9487	0.202 ± 0.462	0.895
	Dataset	—	—	—	0.816
LogP = 0.8	NAT GraphVAE	0.515	1837	0.588 ± 0.759	0.807
	MGM	0.418	9360	0.769 ± 0.473	0.826
	MGM - Final Step	0.300	9294	0.745 ± 0.442	0.857
	Dataset	—	—	—	0.797

The results shown here correspond to the best mean property value (MGM) or the final sampling iteration with initialization chosen according to the better geometric mean among the five GuacaMol metrics (MGM—Final Step). Results for the NAT GraphVAE baseline model²⁵ that we trained are also shown. ‘Dataset’ rows refer to molecules sampled from the dataset with MolWt within ± 1 for the MolWt conditions and LogP within ± 0.1 for the LogP conditions. G-mean refers to the geometric mean of validity, uniqueness and novelty.