Generative Deep Learning for Targeted Compound Design.

Area of Science:

• Computational chemistry
• Artificial intelligence in chemistry
• Molecular modeling

Background:

• *De novo* molecular design utilizes generative models from Deep Learning.
• Applications span drug discovery, materials science, and biotechnology.
• Deep generative models (e.g., RNNs, Autoencoders, GANs) learn from data to propose novel compounds.

Purpose of the Study:

• To systematically review deep generative models for targeted compound design.
• To critically analyze related optimization methods.
• To discuss applications in various scientific fields.

Main Methods:

• Review of deep generative model architectures (RNNs, Autoencoders, GANs).
• Analysis of optimization strategies (transfer learning, Bayesian optimization, reinforcement learning, conditional generation).
• Examination of model training on existing datasets and property distribution adherence.

Main Results:

• Deep generative models can generate novel compounds with desired properties.
• Optimization strategies effectively direct compound generation towards specific aims (biological activity, synthesis, chemical features).
• These models offer a powerful approach to *de novo* molecular design.

Conclusions:

• Deep generative models and optimization techniques are revolutionizing targeted compound design.
• This review provides a critical overview of current methodologies and applications.
• The field holds significant promise for accelerating scientific discovery.