PropMolFlow: property-guided molecule generation with geometry-complete flow matching

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning for chemistry

Background:

• Flow-matching methods achieve state-of-the-art (SOTA) in unconditional molecule generation.
• Diffusion models currently lead in property-guided molecule generation.
• Existing methods face challenges in generating molecules with specific desired properties efficiently and stably.

Purpose of the Study:

• To introduce PropMolFlow, a novel approach for property-guided molecule generation.
• To leverage geometry-complete SE(3)-equivariant flow matching for enhanced molecular design.
• To improve upon SOTA diffusion models in conditional molecule generation tasks.

Main Methods:

• Developed PropMolFlow based on SE(3)-equivariant flow matching.
• Integrated five distinct property embedding techniques with Gaussian expansion for scalar properties.
• Evaluated performance against SOTA diffusion models in conditional generation.

Main Results:

• PropMolFlow demonstrates competitive performance in conditional molecule generation.
• Achieved high structural stability and validity of generated molecules.
• Showcased significantly faster sampling speeds with fewer time steps compared to baseline models.
• Introduced a new task to assess out-of-distribution generalization for under-represented properties.

Conclusions:

• PropMolFlow offers a powerful and efficient alternative for property-guided molecule generation.
• The method maintains high quality and stability in generated molecular structures.
• Validated the importance of density functional theory (DFT) calculations for property verification.
• PropMolFlow shows promise for exploring novel chemical spaces and designing molecules with specific functionalities.