VFMol: A Discrete Flow Matching Variational Autoencoder for Molecular Graph Generation.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Molecular graph generation is crucial for identifying novel drug candidates.
• Variational autoencoders (VAEs) and discrete flow matching (DFM) are used but have limitations.
• Existing methods struggle with permutation invariance, generation bottlenecks, and adaptable prior initialization.

Purpose of the Study:

• To introduce VFMol, a novel framework integrating VAEs and DFM for improved molecular generation.
• To address limitations of VAE decoders and DFM prior initialization.
• To enable efficient, controllable, and high-quality molecular graph generation.

Main Methods:

• VFMol synergistically combines personalized VAE latent space modeling with DFM's stepwise sampling.
• The encoder learns a tailored posterior distribution for each input graph.
• A property-guided framework using KAN and classifier-free guidance enables conditional generation.

Main Results:

• VFMol achieves state-of-the-art performance on molecular generation tasks.
• Demonstrates superior molecular structural quality and property controllability.
• Validates the framework's generality and effectiveness across datasets.

Conclusions:

• VFMol offers a significant advancement in molecular graph generation for drug discovery.
• The integrated approach overcomes key limitations of prior methods.
• VFMol provides a powerful tool for designing novel compounds with desired properties.