Table 2. Results on the Three Molecule Generation Tasks, Each Rerun for Four Different Proteins, Composing Our Docking-Based Benchmarka.
| 5HT1B | 5HT2B | ACM2 | CYP2D6 | |||||
|---|---|---|---|---|---|---|---|---|
| (a) Docking Score Function (↓) | ||||||||
| CVAE | –4.647 | (0.907) | –4.188 | (0.913) | –4.836 | (0.905) | - | - |
| GVAE | –4.955 | (0.901) | –4.641 | (0.887) | –5.422 | (0.898) | - | - |
| REINVENT | -9.774 | (0.506) | -8.657 | (0.455) | -9.775 | (0.467) | -8.759 | (0.626) |
| Train (50%) | –8.541 | (0.850) | –7.709 | (0.878) | –6.983 | (0.868) | –6.492 | (0.897) |
| Train (10%) | –10.837 | (0.749) | –9.769 | (0.831) | –8.976 | (0.812) | –9.256 | (0.869) |
| Train (1%) | –11.493 | (0.859) | –10.023 | (0.746) | –10.003 | (0.773) | –10.131 | (0.763) |
| ZINC (50%) | –7.886 | (0.884) | –7.350 | (0.879) | –6.793 | (0.873) | –6.240 | (0.883) |
| ZINC (10%) | –9.894 | (0.862) | –9.228 | (0.851) | –8.282 | (0.860) | –8.787 | (0.853) |
| ZINC (1%) | –10.496 | (0.861) | –9.833 | (0.838) | –8.802 | (0.840) | –9.291 | (0.894) |
| (b) Repulsion (↓) | ||||||||
| CVAE | 1.148 | (0.919) | 1.001 | (0.914) | 1.132 | (0.908) | 2.234 | (0.914) |
| GVAE | 1.361 | (0.910) | 1.159 | (0.942) | 1.383 | (0.917) | - | - |
| REINVENT | 1.544 | (0.811) | 1.874 | (0.859) | 2.262 | (0.845) | 2.993 | (0.858) |
| Train (50%) | 2.099 | (0.845) | 1.792 | (0.881) | 1.434 | (0.863) | 6.508 | (0.895) |
| Train (10%) | 0.835 | (0.863) | 0.902 | (0.893) | 0.779 | (0.888) | 2.823 | (0.904) |
| Train (1%) | 0.550 | (0.858) | 0.621 | (0.963) | 0.553 | (0.921) | 1.284 | (0.956) |
| ZINC (50%) | 1.803 | (0.878) | 1.677 | (0.882) | 1.665 | (0.879) | 5.786 | (0.880) |
| ZINC (10%) | 0.840 | (0.880) | 0.865 | (0.896) | 0.792 | (0.881) | 2.348 | (0.887) |
| ZINC (1%) | 0.613 | (0.941) | 0.625 | (0.922) | 0.612 | (0.938) | 1.821 | (0.880) |
| (c) Hydrogen Bonding (↑) | ||||||||
| CVAE | 1.089 | (0.915) | 1.168 | (0.909) | 0.881 | (0.907) | 0.539 | (0.908) |
| GVAE | 4.152 | (0.921) | 2.954 | (0.912) | 2.567 | (0.927) | 2.732 | (0.902) |
| REINVENT | 3.795 | (0.626) | 2.451 | (0.580) | 3.520 | (0.480) | 1.304 | (0.574) |
| Train (50%) | 1.069 | (0.843) | 0.668 | (0.882) | 0.296 | (0.871) | 0.684 | (0.892) |
| Train (10%) | 2.934 | (0.751) | 2.327 | (0.816) | 1.444 | (0.896) | 2.061 | (0.884) |
| Train (1%) | 3.351 | (0.825) | 3.586 | (0.575) | 2.519 | (0.852) | 2.700 | (0.917) |
| ZINC (50%) | 1.114 | (0.879) | 0.871 | (0.882) | 0.512 | (0.877) | 0.660 | (0.877) |
| ZINC (10%) | 3.623 | (0.873) | 2.674 | (0.887) | 2.449 | (0.874) | 1.831 | (0.878) |
| ZINC (1%) | 5.743 | (0.928) | 3.545 | (0.935) | 3.253 | (0.940) | 2.115 | (0.861) |
a
The key task is Docking Score Function in which the goal is to optimize the docking score against a given drug target. Each cell reports the mean score for 250 generated molecules in each task. In the parentheses, the internal diversity of generated molecules is reported (see text for details). The tested models tend to improve upon the mean score in the ZINC database (ZINC). However, they generally do not improve upon the top molecules from ZINC; ZINC (10%) and ZINC (1%) show the top 10% of scores and the top 1% of scores. Missing results (“-”) indicate that the model failed to generate 250 molecules that satisfy drug-like filters (see text for details).