Table 4. Average ROC-AUC and Enrichment Statistics across All High-Throughput Screening Assays in PubChemHTSa.
| ROC-AUC | 0.5% | 1% | 2% | |
|---|---|---|---|---|
| Tanimoto similarity | 0.51 | 2.0 | 1.6 | 1.3 |
| MolBERT + ANP | 0.53 | 1.9 | 1.9 | 1.6 |
| NGGP | 0.49 | 1.6 | 1.5 | 1.5 |
| MetaNet | 0.50 | 1.1 | 1.0 | 1.2 |
| Meta-MGNN | 0.53 | 1.5 | 1.5 | 1.4 |
| MAML | 0.54 | 1.3 | 1.2 | 1.3 |
| MetaDTA | 0.52 | 1.9 | 1.5 | 1.4 |
| FS-CAP | 0.56 | 2.3 | 2.1 | 2.0 |
a
Enrichment is shown as a ratio, where 1 means no enrichment over the base rate. Eight context compounds were used for the few-shot learning methods. We ran three replicate training runs with different random number seeds for the three best performing methods (MAML, MetaDTA, FS-CAP) and obtained standard deviations in ROC-AUC of at most 0.01.