GPT-Style Drug Design Using Electron Density from Cryo-EM and X-ray

A new generative model, EDMolGPT, uses low-resolution electron density (ED) from cryo-EM or X-ray experiments as a physical condition for structure-based drug design. Unlike prior methods that condition on empty binding pockets, EDMolGPT leverages ED derived from the filler (ligands and solvent) to capture conformational flexibility and provide a more faithful binding environment description. The framework is a decoder-only autoregressive model that generates molecules from ED point clouds, supporting unified pre-training and experimental integration. The approach was introduced in a preprint on arXiv (2605.08767v1).