A de novo molecular generation method using latent vector based generative adversarial network.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery

Background:

• Deep learning models are increasingly used for generating novel molecular structures.
• Existing generative models may have limitations in exploring diverse chemical spaces.

Purpose of the Study:

• To introduce LatentGAN, a novel deep learning architecture for de novo molecular design.
• To evaluate LatentGAN's performance in generating random and target-biased drug-like compounds.

Main Methods:

• Developed LatentGAN, integrating an autoencoder and a generative adversarial network.
• Applied LatentGAN to generate random and target-biased chemical compounds.
• Compared LatentGAN with a Recurrent Neural Network-based generative model.

Main Results:

• LatentGAN successfully generated random and target-biased drug-like compounds.
• Generated compounds occupied similar chemical space to the training set while introducing novel structures.
• Drug-likeness scores of generated compounds were comparable to the training set.
• LatentGAN produced distinct compounds compared to the Recurrent Neural Network approach.

Conclusions:

• LatentGAN is an effective deep learning architecture for de novo molecular design.
• The method generates novel, drug-like compounds with high fidelity to the training data.
• LatentGAN and Recurrent Neural Network models offer complementary approaches for molecular generation.