Comprehensive Benchmark Study of Diffusion-Based 3D Molecular Generation Models

Deep diffusion models have emerged as powerful tools for de novo molecular generation in three-dimensional (3D) conformational space, leveraging forward noising and reverse denoising processes to learn the probability distributions of molecular geometries and their physicochemical properties. However, systematic comparisons of their performance remain limited. Here, we present a comprehensive benchmark of nine state-of-the-art diffusion-based 3D molecular generative models trained on two representative data sets, QM9 and GEOM-Drugs. Evaluation was performed using four types of metrics, encompassing both two-dimensional (2D) structural features and 3D geometric characteristics. Our results demonstrate that nearly all models perform worse on 3D metrics compared to 2D metrics. Most generated 3D structures exhibit significant deviations from the energy-minimized reference, highlighting the persistent challenges in achieving accurate 3D spatial modeling. Furthermore, their effectiveness declines when producing 3D structures of larger and more complex molecules. Among the models, MiDi and EQGAT-diff consistently outperform the other models, with MiDi showing particularly robust performance. In summary, this study highlights recent advances in deep diffusion models for 3D molecular generation while delineating the key challenges that need to be addressed by next-generation approaches.