MolGuidance: Advanced Guidance Strategies for Conditional Molecular Generation with Flow Matching

Area of Science:

• Computational Chemistry
• Machine Learning
• Drug Discovery

Background:

• Conditional molecular generation aims for chemical validity, property alignment, diversity, and efficient sampling.
• Computer vision guidance strategies offer potential for enhancing generative models.

Purpose of the Study:

• Integrate state-of-the-art guidance methods into a molecular generation framework.
• Develop a hybrid guidance strategy for continuous and discrete molecular modalities.
• Optimize guidance scales using Bayesian optimization.

Main Methods:

• Implemented classifier-free guidance, autoguidance, and model guidance within an SE(3)-equivariant flow matching process.
• Proposed a hybrid strategy guiding velocity fields and denoising probabilities separately.
• Utilized Bayesian optimization for joint guidance scale optimization.

Main Results:

• Achieved new state-of-the-art performance in property alignment for de novo molecular generation on QM9 and QMe14S datasets.
• Generated molecules demonstrated high structural validity.
• Systematically compared guidance methods, providing insights into their strengths and limitations.

Conclusions:

• The hybrid guidance strategy significantly improves de novo molecular generation.
• The framework offers a robust approach for property-guided molecule design.
• This work advances the application of guidance techniques in cheminformatics and drug discovery.