Table 2. Comparing Different Zero- and Few-Shot Model Performances Across Different Metrics on FS-Molb.
| Model | Mode | PR AUC (%) | ΔPR AUC (%) |
|---|---|---|---|
| TwinBooster | zero-shot | 68.56± 0.24 | 20.84±0.24 |
| CLAMPa,(12) | zero-shot | 66.55 ± 0.20 | 19.37 ± 0.20 |
| PN(20) | 16-shot | 67.72 ± 1.00 | 20.17 ± 1.00 |
| GNN-MAML(20) | 16-shot | 63.04 ± 0.96 | 15.49 ± 0.96 |
| GNN-MT(20) | 16-shot | 58.73 ± 0.67 | 11.17 ± 0.67 |
| RF(20) | 16-shot | 56.40 ± 0.78 | 8.85 ± 0.78 |
| MAT(20) | 16-shot | 53.25 ± 0.55 | 5.70 ± 0.55 |
| GNN-ST(20) | 16-shot | 49.61 ± 0.55 | 2.06 ± 0.55 |
a
Only means and standard deviations are given, not by task performance.
b
In zero-shot mode, no “test” molecules are provided; in the case of the few-shot performance, 16 molecules of the “test” set are provided. 10 replicates are carried out for each task; the mean and standard deviation of means are reported. Results that are both the best and statistically significant (Wilcoxon test44 or Welch’s t-test44 in the case of CLAMP, α = 0.05 with the Benjamini-Hochberg45 multiple test correction) are highlighted in bold.