Directly optimizing for synthesizability in generative molecular design using retrosynthesis models.

Area of Science:

• Computational chemistry
• Artificial intelligence in chemistry
• Molecular design

Background:

• Synthesizability is a key challenge in generative molecular design.
• Current methods rely on heuristics or post-hoc retrosynthesis analysis.
• Integrating synthesizability prediction directly into generation is difficult due to computational costs.

Purpose of the Study:

• To develop a method for directly optimizing molecular synthesizability within a generative model.
• To explore the use of retrosynthesis models in goal-directed molecular generation.
• To compare the effectiveness of direct retrosynthesis integration versus heuristic-based approaches.

Main Methods:

• Employed a sample-efficient generative model capable of direct optimization.
• Integrated retrosynthesis models into the goal-directed generation loop.
• Evaluated performance under constrained computational budgets for drug discovery tasks.

Main Results:

• Successfully generated synthesizable molecules satisfying multi-parameter optimization criteria.
• Found that common synthesizability heuristics correlate with retrosynthesis solvability for bioactive molecules.
• Demonstrated diminished correlation of heuristics for functional materials, highlighting the advantage of direct retrosynthesis integration.
• Showcased how over-reliance on heuristics can overlook promising molecules.

Conclusions:

• Directly optimizing for synthesizability using retrosynthesis models is feasible with efficient generative approaches.
• Retrosynthesis models offer advantages over heuristics for diverse molecular classes, especially functional materials.
• This approach enhances the discovery of novel, synthesizable molecules for various applications.