Table 1 Combined results for all compounds in the TTD database for beam size B = 1000 and the different transformer models

TASK	VALIDITY	UNIQUENESS			TOP IDENTICAL			RANK SCORE	CORRELATION
		P	C	NS	P	C	NS
\({\mathcal{D}}\), λ = 0	1.00	1.00	0.97	0.60	0.04	0.04	0.62	0.24 ± 0.27	0.37 ± 0.13
\({\mathcal{D}}\), λ = 10	0.99	1.00	0.95	0.53	0.31	0.30	0.93	0.35 ± 0.25	0.56 ± 0.17
\({{\mathcal{D}}}^{c}\), λ = 0	1.00	1.00	0.97	0.59	0.06	0.07	0.66	0.29 ± 0.25	0.39 ± 0.14
\({{\mathcal{D}}}^{c}\), λ = 10	0.99	1.00	0.95	0.53	0.31	0.31	0.93	0.44 ± 0.24	0.60 ± 0.19

A higher value is the better for all the columns, and the best results are highlighted in bold. \({\mathcal{D}}\) and \({{\mathcal{D}}}^{c}\) represent the training sets generated with ECFP4 fingerprints without and with counts, respectively. The sub-columns P, C, and NS under UNIQUENESS and TOP IDENTICAL denotes different type of post-processing applied to the generated target compounds. The sub-columns are fraction unique SMILES strings (P), fraction unique SMILES strings after canonicalization (C), and fraction unique SMILES strings after removing stereo-chemical information and canonicalization (NS). λ = 0 denotes the absence and λ = 10 the presence of the regularization term when training the transformer models. Best results are highlighted in bold.

Quick links

Search