Drop a molecular library here
CSV / TSV with molecular strings, or SDF / MOL. Up to 50,000 molecules · 64 MB.
Recent tasks
Preview and SMILES column
Checks the first five rows. Invalid molecules remain in the output with an explanation.
| # | SMILES | Parse check |
|---|
| # | Molecule | Applicability | Seed spread | Inspect | ||
|---|---|---|---|---|---|---|
| Start screening to view ranked molecules. | ||||||
Screening guidance and output fields
Where to truncate
this modelrandom expectationtop 5% cut (k=1,108)
Figure 1 The solid line shows model hits; the dashed line shows expected hits from random selection at the same budget. The gap narrows as k increases; both lines meet at the full library.
What this threshold means overall
By score band
| band | entities | positive rate |
|---|---|---|
| high score ≥ 0.634 | — | — |
| grey band 0.534–0.634 | — | — |
| low < 0.534 | — | — |
Confusion matrix at threshold 0.584
| actually positive | actually negative | |
|---|---|---|
| called positive | — | — |
| called negative | — | — |
This threshold favors precision over recall, consistent with candidate screening.
Output format
| column | meaning |
|---|---|
| input columns | Passed through unchanged, same names and same order (whatever the SMILES column is called comes back as it was) |
| connectivity_smiles | The structure the model actually reads, stereo removed |
| score | Score from reference seed 42, 0–1 |
| percentile | Percentile among the 22,168 CSD test entities, 0–1; 0.88 means above 88% of them |
| seed_spread | The gap between the highest and lowest of the three seeds (not a standard deviation); interpret together with applicability and the test-set distribution |
| score_std | Deprecated alias of seed_spread, same value. Kept so older scripts keep reading; will be removed once they have moved over. |
| band | high / gray / low, split at the threshold ±0.05 |
| decision | positive / negative, binarised at 0.584 |
| applicability | in_domain / borderline / out_of_domain, banded by the next column (unknown elements and tiny molecules are handled separately) |
| domain_distance | PCA-whitened distance from the descriptor-space centre of the training set; the training median is about 0.8 |
| domain_percentile | Where that distance falls in the training distribution, 0–1; above 0.95 is borderline, above 0.995 is out_of_domain |
| model_version | v10.2026-10-22 |
| error | Rows that failed to parse explain themselves here; valid rows are empty |
Progress updates automatically after submission. Download the results when calculation finishes.