FlowMol3: flow matching for 3D de novo small-molecule generation

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• Generative models are crucial for accelerating chemical discovery by designing molecules with specific properties.
• Current models often struggle to simultaneously generate molecular topology and 3D structure accurately.
• Advancing generative models for all-atom, small-molecule generation is a key objective in computational chemistry.

Purpose of the Study:

• To present FlowMol3, an open-source, multi-modal flow matching model for state-of-the-art small-molecule generation.
• To demonstrate performance gains through architecture-agnostic techniques without altering the core model.
• To improve the stability and quality of transport-based generative models.

Main Methods:

• Developed FlowMol3, a flow matching model for molecular generation.
• Implemented three architecture-agnostic techniques: self-conditioning, fake atoms, and train-time geometry distortion.
• Evaluated model performance on molecular validity, functional group composition, and geometric accuracy.

Main Results:

• FlowMol3 achieves nearly 100% molecular validity for drug-like molecules with explicit hydrogens.
• The model accurately reproduces functional group composition and geometry of training data.
• FlowMol3 requires an order of magnitude fewer parameters than comparable methods while improving quality.

Conclusions:

• Novel techniques like self-conditioning, fake atoms, and geometry distortion significantly enhance molecular generative models.
• These methods mitigate distribution drift in transport-based models, improving inference stability and quality.
• FlowMol3 offers simple, transferable strategies for advancing diffusion- and flow-based molecular generation.