Skip to content

checkpoints.md

Latest commit

 

History

History
29 lines (21 loc) · 1.97 KB

checkpoints.md

File metadata and controls

29 lines (21 loc) · 1.97 KB

Checkpoints on s1b dataset

Initial dataset used for model validation and optimization. Notice that, given randomness, differences in library versions and GNEprop library itself, results are slightly different compared to those reported in the manuscript, but the overall trends are consistent. 20 folds each.

Checkpoint Comment AUPRC AUROC F1
20250801-153038 GNEprop (no pretraining, no RDKit features) 0.486 0.807 0.456
20250801-154530 GNEprop (no pretraining, + RDKit features) 0.497 0.825 0.433
20250801-150551 GNEprop (+ pretraining, no RDKit features) 0.558 0.857 0.487
20250801-182534 GNEprop (+ pretraining, + RDKit features) 0.560 0.848 0.482

We also release checkpoint 20250817-181637 which includes meta-learning-based fine-tuning on top of the best model (not reported in the manuscript), further improving it (AUPRC = 0.560, AUROC = 0.873, F1 = 0.493)

Checkpoints on GNEtolC dataset

Notice that GNEprop has not been explicitly hyper-optimized or evaluated on this dataset. 8 folds each.

  • 20250819-085119: GNEprop trained on GNEtolC dataset with scaffold splitting
  • 20250819-093608: GNEprop trained on GNEtolC dataset with scaffold-cluster splitting

Self-supervised checkpoint

  • 20210827-082422: self-supervised checkpoint

GNEprop trained on HTS dataset

  • 20250811-202022: GNEprop trained on the full HTS dataset (95/5 random splitting). 8 folds. In addition to self-supervised pre-training (as in the manuscript), this checkpoint also leverages adversarial augmentations (which gave an additional minor improvement on the public dataset). Given randomness and version changes, predictions are slightly different from those reported for the virtual hits, although highly correlated (Spearman correlation = 0.62 on the virtual hits subset).