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 sampling realistic molecules with desired properties.
• Developing models that jointly sample molecular topology and 3D structure is a key challenge in this field.
• Existing methods often require complex architectures or extensive computational resources.

Purpose of the Study:

• To present FlowMol3, an open-source, multi-modal flow matching model for advanced all-atom, small-molecule generation.
• To demonstrate significant performance gains over previous FlowMol versions through architecture-agnostic techniques.
• To improve the stability and quality of transport-based generative models for molecular design.

Main Methods:

• Utilized self-conditioning, fake atoms, and train-time geometry distortion as architecture-agnostic techniques.
• Implemented a multi-modal flow matching approach for joint sampling of molecular topology and 3D structure.
• Focused on achieving high molecular validity and accurate reproduction of training data characteristics.

Main Results:

• FlowMol3 achieves nearly 100% molecular validity for drug-like molecules with explicit hydrogens.
• The model demonstrates more accurate reproduction of functional group composition and geometry compared to previous versions.
• FlowMol3 requires an order of magnitude fewer learnable parameters than comparable methods, indicating high efficiency.

Conclusions:

• The presented techniques effectively mitigate distribution drift in transport-based generative models.
• FlowMol3 offers a stable and high-quality solution for molecular generation, advancing the state of the art.
• These simple, transferable strategies can enhance the performance of diffusion- and flow-based molecular generative models.