Table 3 Performance comparison of the GMT models trained with 20%, 50%, and 100% training samples
From: Probabilistic generative transformer language models for generative design of molecules
Training samples | 20% | 50% | 100% | |
|---|---|---|---|---|
Valid | \(\uparrow\) | 1.0000 | 1.0000 | 1.0000 |
Unique@1000 | \(\uparrow\) | 1.0000 | 1.0000 | 1.0000 |
Unique@10000 | \(\uparrow\) | 1.0000 | 0.9998 | 1.0000 |
FCD/Test | \(\downarrow\) | 4.3961 | 3.9164 | 3.7750 |
SNN/Test | \(\uparrow\) | 0.4526 | 0.4573 | 0.4673 |
Frag/Test | \(\uparrow\) | 0.9840 | 0.9850 | 0.9869 |
Scaf/Test | \(\uparrow\) | 0.8225 | 0.8049 | 0.8431 |
FCD/TestSF | \(\downarrow\) | 5.2401 | 4.7000 | 4.5698 |
SNN/TestSF | \(\uparrow\) | 0.4362 | 0.4395 | 0.4485 |
Frag/TestSF | \(\uparrow\) | 0.9792 | 0.9802 | 0.9831 |
Scaf/TestSF | \(\uparrow\) | 0.1340 | 0.1461 | 0.1096 |
IntDiv | \(\uparrow\) | 0.8707 | 0.8704 | 0.8701 |
IntDiv2 | \(\uparrow\) | 0.8653 | 0.8650 | 0.8646 |
Filters | \(\uparrow\) | 0.7858 | 0.7913 | 0.7961 |
Novelty | \(\uparrow\) | 0.9790 | 0.9751 | 0.9683 |