Table 2 Comparison of PCF-VAE model on MOSES benchmark.
From: PCF-VAE: posterior collapse free variational autoencoder for de novo drug design
Model | Valid (\(\uparrow\)) | Unique@1k (\(\uparrow\)) | Unique@10k (\(\uparrow\)) | IntDiv (\(\uparrow\)) | IntDiv2 (\(\uparrow\)) | Filters (\(\uparrow\)) | Novelty (\(\uparrow\)) |
|---|---|---|---|---|---|---|---|
Train | 1.0 | 1.0 | 1.0 | 0.8567 | 0.8508 | 1.0 | 1.0 |
HMM | 0.076 ± 0.0322 | 0.623 ± 0.1224 | 0.5671 ± 0.1424 | 0.8466 ± 0.0403 | 0.8104 ± 0.0507 | 0.9024 ± 0.0489 | 0.9994 ± 0.001 |
NGram | 0.2376 ± 0.0025 | 0.974 ± 0.0108 | 0.9217 ± 0.0019 | 0.8738 ± 0.0002 | 0.8644 ± 0.0002 | 0.9582 ± 0.001 | 0.9694 ± 0.001 |
Combinatorial | 1.0 ± 0.0 | 0.9983 ± 0.0015 | 0.9909 ± 0.0009 | 0.8732 ± 0.0002 | 0.8666 ± 0.0002 | 0.9557 ± 0.0018 | 0.9878 ± 0.0008 |
CharRNN | 0.9748 ± 0.0264 | 1.0 ± 0.0 | 0.9994 ± 0.0003 | 0.8562 ± 0.0005 | 0.8503 ± 0.0005 | 0.9943 ± 0.0034 | 0.8419 ± 0.0509 |
AAE | 0.9368 ± 0.0341 | 1.0 ± 0.0 | 0.9973 ± 0.002 | 0.8557 ± 0.0031 | 0.8499 ± 0.003 | 0.996 ± 0.0006 | 0.7931 ± 0.0285 |
VAE | 0.9767 ± 0.0012 | 1.0 ± 0.0 | 0.9984 ± 0.0005 | 0.8558 ± 0.0004 | 0.8498 ± 0.0004 | 0.997 ± 0.0002 | 0.6949 ± 0.0069 |
JTN-VAE | 1.0 ± 0.0 | 1.0 ± 0.0 | 0.9996 ± 0.0003 | 0.8551 ± 0.0034 | 0.8493 ± 0.0035 | 0.976 ± 0.0016 | 0.9143 ± 0.0058 |
LatentGAN | 0.8966 ± 0.0029 | 1.0 ± 0.0 | 0.9968 ± 0.0002 | 0.8565 ± 0.0007 | 0.8505 ± 0.0006 | 0.9735 ± 0.0006 | 0.9498 ± 0.0006 |
TGVAE 1H/M | 0.948 ± 0.018 | – | 0.999 ± 0.000 | 0.864 ± 0.003 | 0.861 ± 0.003 | – | 0.964 ± 0.004 |
PCF-VAE, \(D_1\) | 0.9801 ± 0.0013 | 1.0 ± 0.0 | 1.0 ± 0.0 | 0.8587 ± 0.0001 | 0.8527 ± 0.0001 | 0.990 ± 0.0387 | 0.9377 ± 0.0002 |
PCF-VAE, \(D_2\) | 0.9710 ± 0.0011 | 1.0 ± 0.0 | 1.0 ± 0.0 | 0.8881 ± 0.0003 | 0.8621 ± 0.0003 | 0.981 ± 0.0271 | 0.9471 ± 0.0101 |
PCF-VAE, \(D_3\) | 0.9501 ± 0.0101 | 1.0 ± 0.0 | 1.0 ± 0.0 | 0.8901 ± 0.1201 | 0.8633 ± 0.0004 | 0.971 ± 0.0281 | 0.9501 ± 0.0102 |
