Molecular graph generation is a key task in drug discovery, aiming to efficiently identify novel compounds with desired properties. While variational autoencoders (VAEs) excel at latent space modeling and discrete flow matching (DFM) enables efficient continuous-time sampling, existing approaches still face critical limitations. VAE decoders struggle with permutation invariance and suffer from one-shot generation bottlenecks, whereas DFM models often rely on a fixed prior initialization that lacks adaptability to specific molecular structures. To address these issues, we propose VFMol, a novel framework that synergistically integrates personalized VAE latent space modeling with the efficient stepwise sampling mechanism of DFM in the discrete space. Specifically, the encoder learns a posterior distribution tailored to each input graph as the generation starting point, thereby enhancing both structural fidelity and diversity. Moreover, we introduce a lightweight property-guided framework based on KAN and classifier-free guidance, enabling conditional generation without auxiliary property predictors. Experiments on two widely used molecular data sets demonstrate that VFMol achieves state-of-the-art performance in terms of molecular structural quality and property controllability, verifying its generality and effectiveness.
Gan et al. (Thu,) studied this question.