Communications Chemistry

Table 1 Comparison of distribution learning metrics for each method across five benchmark datasets, MOSES, GuacaMol, Polymer, SuperNatural3, and ZINC250K

From: Leveraging tree-transformer VAE with fragment tokenization for high-performance large chemical model generation

MOSES dataset
Model	Recon	Similar	Valid	Unique	Novelty	FCD	KL-Div.	logP	QED	NP	SA
FRATTVAE	0.9487	0.9847	1.0000	0.9997	0.9742	0.8654	0.9281	0.2032	0.0303	0.1853	2.591
JTVAE*	0.5109	0.7418	0.9989	0.9997	0.9511	0.7933	0.9632	0.1319	0.0043	0.2954	2.728
PSVAE	0.0250	0.2306	1.0000	0.9961	0.9898	0.2259	0.8314	0.4720	0.0966	0.9996	3.488
MoLeR	0.1650	0.5262	1.0000	0.9997	0.9710	0.8525	0.9533	0.0944	0.0270	0.1622	2.485
SMIVAE	0.0000	0.1287	0.2192	0.0838	1.0000	0.0000	0.1503	53.8212	0.7506	1.6976	4.618
SMITransVAE	0.9920	0.9946	0.2635	0.9435	0.9955	0.0739	0.8122	0.6934	0.0775	0.7607	2.570

GuacaMol dataset
FRATTVAE	0.7606	0.8976	1.0000	0.9997	0.9674	0.8242	0.9526	0.0905	0.0147	0.0965	2.981
PSVAE	0.0335	0.1503	0.9999	0.9955	0.9893	0.2049	0.8110	0.8438	0.0702	0.5813	3.886
MoLeR*	0.0252	0.3047	1.0000	0.9995	0.9900	0.6413	0.9638	0.1191	0.0353	0.2006	2.777
SMIVAE	0.0000	0.0675	0.9967	0.0572	0.9883	0.0000	0.2349	2.6651	0.3963	0.9588	3.731
SMITransVAE	0.9801	0.9865	0.4465	0.2686	0.9639	0.0008	0.5853	4.1386	0.1753	1.3113	2.978

Polymer dataset
FRATTVAE	0.9766	0.9957	1.0000	0.9937	0.8782	0.9002	0.9571	0.2986	0.0085	0.0241	4.371
JTVAE	0.4720	0.8182	0.9821	0.9743	0.8871	0.5150	0.9266	0.9349	0.0226	0.1668	5.156
HierVAE*	0.7994	0.9487	1.0000	0.9644	0.6278	0.9101	0.9893	0.8820	0.0123	0.0245	4.306
PSVAE	0.0006	0.1601	0.9973	0.6639	0.9900	0.0940	0.7630	4.8181	0.1352	0.1887	4.924
MoLeR	0.4086	0.7787	1.0000	0.9325	0.5710	0.9019	0.9815	0.4652	0.0211	0.0377	4.190
SMIVAE	0.8010	0.9462	1.0000	0.9452	0.8838	0.7836	0.9226	0.8143	0.0231	0.0792	4.146
SMITransVAE	0.0006	0.2978	0.9674	0.0503	0.3380	0.2274	0.9061	4.5155	0.0433	0.0505	4.512

Natural Products dataset (SuperNatural3)
FRATTVAE	0.3221	0.7774	1.0000	0.9964	0.9273	0.7437	0.9216	0.9149	0.0265	0.1480	5.4715
SMIVAE	0.1532	0.4929	0.6511	0.9983	0.8895	0.7080	0.9195	0.3669	0.0078	0.1837	5.2158

ZINC250K dataset
FRATTVAE	0.7964	0.8944	1.0000	0.9991	0.9945	0.7163	0.9397	0.2979	0.0758	0.2937	3.3009
C-FRATTVAE	0.7940	0.9041	1.0000	0.9999	0.9974	0.7283	0.9685	0.1351	0.0406	0.2426	3.2188
JTVAE*	0.7798	0.8888	1.0000	0.9989	0.9997	0.4401	0.9060	0.6944	0.0312	0.4101	3.4113
HierVAE	0.0000	0.1034	1.0000	0.0350	0.9988	0.0006	0.6232	1.1876	0.1192	0.6086	3.5890
PSVAE*	0.0006	0.1398	1.0000	0.9958	0.9998	0.3053	0.8717	0.3218	0.0417	0.8091	3.9908
MoLeR	0.0444	0.3664	1.0000	0.9998	0.9997	0.7293	0.9872	0.1367	0.0197	0.1442	3.0399
SMIVAE	0.0184	0.3284	0.9815	0.9984	0.9962	0.7003	0.9337	0.0623	0.0329	0.0913	2.7755
SMITransVAE	0.9718	0.9829	0.3638	0.7700	0.9987	0.0580	0.8010	1.1458	0.0718	0.6028	2.3858

Distribution learning metrics comparison for each method trained on five benchmark datasets, MOSES, GuacaMol, Polymer, SuperNatural3, and ZINC250K. Metrics include reconstruction accuracy (Recon), similarity (Similar), validity (Valid), uniqueness (Unique), novelty (Novelty), Fréchet ChemNet Distance (FCD), KL divergence (KL-Div.), logP, QED, NP (NP-likeness), and SA (SAscore). For three metrics (logP, QED, and NP-likeness), we used the 1D Wasserstein-1 distance between the metric distributions of the generated and test sets, rather than their absolute values. Models marked with an ‘*’ indicate the use of pretrained models.
JTVAE was unable to handle GuacaMol and Natural Products datasets. HierVAE was unable to handle MOSES, GuacaMol and Natural Products datasets. PSVAE and MoLeR were unable to handle Natural Products dataset. SMITransVAE was unable to handle Polymer and Natural Products datasets.

Back to article page

Search

Advanced search

Quick links