TANGO: direct optimization of constrained synthesizability for generative molecular design

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Chemical synthesis

Background:

• Generative molecular design faces challenges in ensuring synthesizability.
• Multi-parameter optimization of molecules often neglects practical synthesis constraints.
• Efficient molecule repurposing, sustainability, and synthesis efficiency are critical.

Purpose of the Study:

• To address the challenge of constrained synthesizability in generative molecular design.
• To develop a novel reward function for optimizing molecule synthesis.
• To integrate synthesizability into generative models using reinforcement learning.

Main Methods:

• Introduction of the Tanimoto Group Overlap (TANGO) reward function.
• Transformation of binary rewards into continuous rewards using chemical principles.
• Augmentation of molecular generative models with TANGO for reinforcement learning (RL).
• Framework designed to handle various synthesis constraints (starting-material, intermediate, divergent).

Main Results:

• TANGO enables direct optimization for constrained synthesizability in generative models.
• The proposed method successfully integrates synthesis planning into molecular design.
• Demonstrated effectiveness in navigating complex synthesis optimization scenarios.
• RL-based incentivization proves productive for challenging synthesizability problems.

Conclusions:

• The TANGO reward function offers a novel approach to optimizing molecular synthesizability.
• Reinforcement learning with TANGO enhances generative models for practical chemical design.
• The framework provides a general solution for incorporating diverse synthesis constraints.