Molecular optimization using a conditional transformer for reaction-aware compound exploration with reinforcement...

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Deep learning advances have enabled molecular generative models for drug discovery.
• Existing models often overlook the crucial aspect of synthetic feasibility.
• Navigating vast chemical spaces requires consideration of real-world reactivity constraints.

Purpose of the Study:

• To introduce TRACER, a novel framework combining molecular property optimization with synthetic pathway generation.
• To address the limitations of current models in ensuring the practical synthesis of generated molecules.
• To enhance the efficiency and applicability of AI-driven molecular design in drug discovery.

Main Methods:

• TRACER utilizes a conditional transformer model to predict reaction products from reactants.
• The framework integrates molecular property optimization with de novo synthetic pathway generation.
• The model operates under defined reaction type constraints to ensure synthetic feasibility.

Main Results:

• TRACER effectively generated compounds with high scores in molecular optimization tasks.
• Performance was validated using activity prediction models targeting DRD2, AKT1, and CXCR4.
• The transformer model successfully captured the complexities of organic synthesis and chemical space navigation.

Conclusions:

• TRACER offers a robust solution for designing synthetically feasible molecules with desired properties.
• The framework advances AI-driven approaches in drug discovery by bridging molecular design and synthesis planning.
• TRACER facilitates exploration of vast chemical spaces while adhering to practical synthetic constraints.