chromeNLP/openssi-target-based-drug-screening

DrugCLIP epoch 14

Selected fine-tune checkpoint from the paper-scale DrugCLIP run. Held-out scores below are the corrected rescore in this report.

model.pt SHA-256: bb108dcec5608b04eee546a0ed57248ab5f956cb61799a8c5611ab9c8da7c772

Equal-target mean EF at 1% is 16.108.

Benchmark	Targets	ROC-AUC	BEDROC (80.5)	EF 1%
DUD-E	102	0.769	0.295	18.056
LIT-PCBA	15	0.577	0.038	2.861

Random ranking is EF 1.

The checkpoint is fine-tune epoch 14, selected by development BEDROC 0.4288. Fine-tuning stopped at epoch 34 after patience 20. The model has 95,305,418 parameters. Pocket atoms were scored with element tokens, a 511-atom test crop, and the author pooling (DUD-E median/MAD, then max; LIT-PCBA max).

Exact benchmark means from that rescore:

Benchmark	ROC-AUC	BEDROC (80.5)	EF 1%
DUD-E	0.7686053801137084	0.29458757122655027	18.055988235981932
LIT-PCBA	0.5774894830416493	0.0381821961785978	2.861166091212881

Overall equal-target mean EF at 1% is 16.107934114857695.