Growing and linking optimizers: synthesis-driven molecule design.

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in chemistry

Background:

• Generative models are crucial for designing novel molecules with desired properties.
• Existing models like text-based and graph-based approaches have limitations in emulating real chemical synthesis.
• Drug design requires molecules that are not only effective but also synthetically accessible.

Purpose of the Study:

• To introduce two new reaction-based generative models: growing optimizer and linking optimizer.
• To develop models that emulate chemical synthesis for designing feasible molecules.
• To improve upon existing generative models for drug discovery applications.

Main Methods:

• Developing two reaction-based generative models: growing optimizer and linking optimizer.
• Simulating chemical synthesis by sequentially selecting building blocks and reaction types.
• Incorporating comprehensive chemical knowledge into the generative process.
• Comparing model performance against REINVENT 4, a state-of-the-art model.

Main Results:

• Growing optimizer and linking optimizer successfully emulate real-life chemical synthesis.
• The models can restrict chemistry to specific building blocks, reaction types, and synthesis pathways.
• Generated molecules exhibit higher synthetic accessibility compared to REINVENT 4.
• The models achieve molecules of interest with desired properties.

Conclusions:

• Growing optimizer and linking optimizer represent a significant advancement in reaction-based molecular design.
• These models offer a more chemically informed approach to generative chemistry, crucial for drug discovery.
• The focus on synthetic feasibility enhances the practical utility of generated molecules in pharmaceutical research.