Table 1.
Relevant studies on the GAN-based structures of molecular de novo design.
| Study | Structure | Architecture | Object Generated | Learning Technique | Databases | Results |
|---|---|---|---|---|---|---|
| Kadurin et al. [28,29] | druGAN | AAE | latent vector | autoencoder | PubChem | druGAN generated novel molecular compounds which can be considered as potential anticancer agents. |
| Guimaraes et al. [36] | ORGAN | GAN | SMILES | RL | ZINC, GDB-17 |
ORGAN performed better than recurrent neural networks or GAN alone. |
| Sanchez-Lengeling et al. [37] | ORGANIC | GAN | SMILES | RL | ZINC, GDB-17 |
ORGANIC showed good performance in terms of the quantitative estimate of drug-likeness, but not the Lipinski’s Rule-of-Five. |
| Putin et al. [38] | RANC | GAN | SMILES | RL | ZINC, ChemDiv | RANC was superior to ORGANIC in terms of several drug discovery metrics. |
| Putin et al. [39] | ATNC | GAN | SMILES | RL | ChemDiv | ATNC performed better than ORGANIC in terms of various functions. |
| Polykovskiy et al. [40] | ECAAE | AAE | latent vector | autoencoder | ZINC | ECAAE generated novel molecular compounds which can be considered as target drugs in rheumatoid arthritis, psoriasis, and vitiligo. |
| Cao and Kipf [41] | MolGAN | GAN | graph | RL | QM9 | MolGAN outperformed ORGAN and variational autoencoder-based structures. |
| Guarino et al. [42] | DiPol-GAN | GAN | graph | RL | QM9 | DiPol-GAN had 1.3 times higher drug-likeliness scores than MolGAN. |
| Prykhodko et al. [43] | LatentGAN | GAN | SMILES | autoencoder | ChEMBL | LatentGAN created novel drug-like compounds and was compatible to recurrent neural networks. |
| Maziarka et al. [44] | Mol-CycleGAN | GAN | latent vector | direct flow | ZINC, ChEMBL | Mol-CycleGAN outperformed the junction tree variational autoencoder and the graph convolutional policy network structures. |
| Méndez-Lucio et al. [45] | Conditioned GAN | GAN | latent vector | direct flow | L1000 | Conditioned GAN produced molecular compounds with desired gene expression signatures. |
AAE = adversarial autoencoder; ATNC = Adversarial Threshold Neural Computer; druGAN = drug Generative Adversarial Network; ECAAE = Entangled Conditional Adversarial AutoEncoder; GAN = Generative Adversarial Network; LatentGAN = Latent Generative Adversarial Networks; MolGAN = Molecular Generative Adversarial Network; Mol-CycleGAN = Molecular Cycle Generative Adversarial Network; ORGAN = Objective-Reinforced Generative Adversarial Networks; ORGANIC = Objective-Reinforced Generative Adversarial Network for Inverse-design Chemistry; RANC = Reinforced Adversarial Neural Computer.