Table 1.

AUROC values for the different models trained on a small dataset of 10,000 compounds

Similarity threshold	Vanilla transformer	Triplet loss	Similarity loss
0.45	0.68 ± 0.17	0.73 ± 0.17	0.82 ± 0.18
0.50	0.69 ± 0.18	0.75 ± 0.16	0.86 ± 0.17
0.55	0.75 ± 0.18	0.80 ± 0.15	0.92 ± 0.08
0.60	0.76 ± 0.18	0.81 ± 0.15	0.91 ± 0.11
0.65	0.80 ± 0.17	0.85 ± 0.13	0.94 ± 0.09
0.70	0.84 ± 0.18	0.89 ± 0.12	0.96 ± 0.07
0.75	0.87 ± 0.16	0.91 ± 0.12	0.97 ± 0.07
0.80	0.90 ± 0.14	0.94 ± 0.09	0.98 ± 0.07
0.85	0.92 ± 0.14	0.96 ± 0.08	0.98 ± 0.07
0.90	0.94 ± 0.14	0.98 ± 0.05	0.98 ± 0.08
0.95	0.97 ± 0.09	0.99 ± 0.04	1.00 ± 0.01

While the vanilla transformer model was trained using only a reconstruction loss function, the other two models were trained with an additional loss term to specifically enforce the conservation of ground truth similarities in the latent space