Sampling 3D Molecular Conformers with Diffusion Transformers

The recent introduction of the DiTMC framework marks a significant advancement in the application of Diffusion Transformers (DiTs) for molecular conformer generation. DiTs have shown strong performance in generative modeling, particularly in image synthesis, but applying them to molecular structures presents unique challenges, such as the integration of discrete molecular graph information with continuous 3D geometry. DiTMC addresses these challenges through a modular architecture that separates the processing of 3D coordinates from atomic connectivity conditioning. By employing two complementary graph-based conditioning strategies and various attention mechanisms, DiTMC achieves a balance between accuracy and computational efficiency. Experiments conducted on standard conformer generation benchmarks, including GEOM-QM9, DRUGS, and XL, demonstrate that DiTMC achieves state-of-the-art precision and physical validity. This development not only highlights the impact of architectural choic…