DriftingMol: Decoder-Coupled Drift for One-Pass Property-Conditional Molecular Generation
Researchers have developed DriftingMol, a novel two-stage framework for generating molecules with specific properties. This method adapts drifting models to a SELFIES latent molecular space, utilizing the decoder's hidden representation as a drift feature map. The approach achieves improved property-conditional generation with low sampling costs, demonstrating strong correlations on datasets like ZINC250K for properties such as QED. AI
IMPACT Introduces a low-cost mechanism for property-biased molecular generation, potentially accelerating drug discovery and materials science research.