EvoMol: a flexible and interpretable evolutionary algorithm for unbiased de novo molecular generation.

Area of Science:

• Computational chemistry
• Drug discovery
• Materials science

Background:

• Exploring vast chemical space is crucial for discovering novel molecules.
• Existing molecular generation methods often require prior data or lack flexibility.
• Interpretability is key for understanding and improving generative models.

Purpose of the Study:

• To develop a flexible, open-source molecular generator.
• To enable exploration of both known and novel chemical spaces.
• To ensure model interpretability for diagnosis and enhancement.

Main Methods:

• Utilizing an evolutionary algorithm to sequentially build molecular graphs.
• Implementing 7 generic, atomic-level mutations for broad exploration.
• Designing a method independent of starting data for generating unseen compounds.

Main Results:

• Achieved excellent performance on established benchmarks like QED, penalized logP, SAscore, and CLscore.
• Demonstrated flexibility by generating optimized organic molecules for materials science (high HOMO, low LUMO).
• Successfully applied constraints on synthesizability and structural features.

Conclusions:

• EvoMol is a versatile and high-performing molecular generator.
• The method's interpretability allows visualization of the exploration process.
• EvoMol facilitates the discovery of novel molecules for diverse applications.