ED-DiT Uses Electron Density for Transferable Molecular AI

Pretraining methods have shown great promise for developing transferable representations, but their application to electron-density-based molecular learning has been underexplored. Electron density offers a comprehensive 3D description of a molecule's electronic structure, capturing both local and global physical properties. This raises the question of whether electron-density fields can be used for self-supervised pretraining to create a shared representation applicable to diverse electronic-structure tasks. Researchers propose ED-DiT, a physics-guided Diffusion Transformer designed for self-supervised pretraining on electron-density point clouds. ED-DiT learns robust representations by reconstructing corrupted and partially masked log-density fields across different diffusion noise levels, incorporating an electron-number consistency constraint to maintain total electronic mass. The pretrained encoder can then be adapted for tasks such as property prediction, classification, and electron-density retrieval. Experiments on six EDBench tasks demonstrate ED-DiT's superior performance, particularly under limited supervision, outperforming models trained from scratch and existing baselines.