Table 2.
Performance comparison on the MOSES benchmark
| Model | Validity (↑) | Uniqueness@10k (↑) | Novelty (↑) | Filters (↑) | IntDiv (↑) |
|---|---|---|---|---|---|
| CharRNN | 0.9748 | 0.9994 | 0.8419 | 0.9943 | 0.8562 |
| AAE | 0.9368 | 0.9973 | 0.7931 | 0.9960 | 0.8557 |
| VAE | 0.9767 | 0.9984 | 0.6949 | 0.9970 | 0.8558 |
| JTN-VAE | 1.0000 | 0.9996 | 0.9143 | 0.9760 | 0.8551 |
| LatentGAN | 0.8966 | 0.9968 | 0.9498 | 0.9735 | 0.8565 |
| MolGPT | 0.9900 | 0.9986 | 0.8093 | 0.9972 | 0.8526 |
| MDAGS | 0.9556 | 0.9984 | 0.9925 | 0.9948 | 0.8534 |
The bold value in each column corresponds to the optimal value in all comparison models under this indicator.