EDMolGPT: GPT-Style Drug Design Using Electron Density
Researchers have introduced EDMolGPT, a novel generative modeling framework for structure-based drug design (SBDD) that utilizes low-resolution electron density (ED) as a conditioning signal. Unlike existing methods that rely on rigid, empty binding pockets from holo complexes, this approach leverages ED derived from fillers such as ligands and solvents. This method supports both computationally calculated and experimentally obtained data from cryo-EM or X-ray sources, enabling unified pre-training and better integration with experimental workflows. By using ED point clouds, the model naturally captures conformational flexibility and provides a more accurate description of the binding environment, thereby mitigating structural bias. EDMolGPT operates as a decoder-only autoregressive system, generating molecules with valid 3D conformations grounded in physically meaningful density signals. The effectiveness of this approach was verified through evaluations on 101 biological targets, demonstrating significant progress in de novo drug design. This advancement highlights the potential of combining generative AI with physical data to enhance the precision and realism of molecular generation in pharmaceutical research.
Wire timeline
EDMolGPT: GPT-Style Drug Design Using Electron Density
Researchers have introduced EDMolGPT, a novel generative modeling framework for structure-based drug design (SBDD) that utilizes low-resolution electron density (ED) as a conditioning signal. Unlike existing methods that rely on rigid, empty binding pockets from holo complexes, this approach leverages ED derived from fillers such as ligands and solvents. This method supports both computationally calculated and experimentally obtained data from cryo-EM or X-ray sources, enabling unified pre-training and better integration with experimental workflows. By using ED point clouds, the model naturally captures conformational flexibility and provides a more accurate description of the binding environment, thereby mitigating structural bias. EDMolGPT operates as a decoder-only autoregressive system, generating molecules with valid 3D conformations grounded in physically meaningful density signals. The effectiveness of this approach was verified through evaluations on 101 biological targets, demonstrating significant progress in de novo drug design. This advancement highlights the potential of combining generative AI with physical data to enhance the precision and realism of molecular generation in pharmaceutical research.
cs.AI updates on arXiv.org