عن المشروع

UMADock provides a workflow for docking small molecules into protein binding sites and scoring the resulting poses with pluggable machine-learning interatomic potentials. The README describes a pipeline that creates ligand conformations with RDKit, prepares a binding site from a PDB structure plus a ligand residue name, docks conformers into that site, optimizes promising poses, and combines interaction, desolvation, and strain terms to estimate an electronic binding energy. Site preparation is automated with PDBFixer, residue selection around the ligand, ACE/NME capping, AMBER ff14SB protonation and charges, and optional handling of structural metal ions. The default scorer is Meta’s UMA model, while MACE-OMOL-0 and AIMNet2 are also supported through a selectable scoring model. The project emphasizes running end-to-end on Modal cloud GPUs, with caching for PDBs, prepared sites, and model weights, plus retrieval of optimized poses and result JSON files from Modal Volumes. It also includes a CLI version, example notebooks, and an AI-agent notebook for driving UMADock from an agent. Validated examples include paracetamol and salbutamol in SULT1A3, with comparisons to explicit-solvent MD and MM-GBSA results.