New TopVAE Model Improves 3D Molecular Generation by Reducing "Dark Areas".

Researchers have identified a significant challenge in 3D molecular generation using latent diffusion frameworks: the presence of "dark areas" within the latent space. These regions, while reachable during diffusion sampling, decode into structurally invalid or chemically disconnected molecules, posing a major hurdle for precise molecular design. To tackle this, the paper proposes TopVAE, a topology-optimized Variational Autoencoder. Unlike traditional VAEs that rely on reconstruction objectives, TopVAE is designed to internalize structural and chemical constraints directly into its decoder during the training phase. This eliminates the need for post-generation chemical correction. Experimental results demonstrate that TopVAE substantially improves off-posterior robustness. When combined with a standard DiT model, it achieves significantly lower FCD-3D scores on benchmarks like QM9 and GEOM-Drugs, and generates a higher proportion of stable and connected molecules in zero-shot scaffold inpainting tasks.