Exploring Discrete Flow Matching for 3D De Novo Molecule Generation

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery

Background:

• Deep generative models accelerate chemical discovery by designing novel molecular structures.
• Flow matching, a generative framework, excels with continuous data but requires adaptation for discrete molecular data.
• Existing discrete flow matching methods have limitations for 3D *de novo* small molecule generation.

Purpose of the Study:

• To benchmark existing discrete flow matching methods for 3D *de novo* small molecule generation.
• To introduce FlowMol-CTMC, an open-source model for enhanced 3D molecular design.
• To propose new metrics for evaluating molecule quality beyond basic constraints.

Main Methods:

• Benchmarking of current discrete flow matching techniques for 3D molecular generation.
• Development and implementation of the FlowMol-CTMC model.
• Introduction of novel quality assessment metrics focusing on higher-order structural motifs.

Main Results:

• FlowMol-CTMC achieves state-of-the-art performance in 3D *de novo* small molecule generation.
• The proposed model utilizes fewer learnable parameters compared to existing methods.
• New metrics revealed that generated molecules, while satisfying basic constraints, may contain unusual functional groups outside the training distribution.

Conclusions:

• Discrete flow matching is a viable framework for 3D *de novo* molecular design.
• FlowMol-CTMC offers an efficient and high-performing solution for generating novel molecules.
• Further evaluation using advanced metrics is crucial for ensuring the chemical validity and utility of generated molecules.