How jssweller/drughive works
Project type and purpose
• DrugHIVE is a Python-based, command‐line-driven machine-learning pipeline (a deep hierarchical variational autoencoder) for structure‐based drug design. It ingests 3D protein–ligand data, trains a generative model, then samples or optimizes novel ligands via docking.
File structure
• Top‐level scripts:
• process_pdbbind_data.py, process_zinc_dataset.py – data preprocessing
• train.py – model training entry point
• generate_molecules.py – sampling new molecules
• generate_optimize.py, ff_optimize.py, dock.py – ligand optimization & docking
• Core library: drughive/
• atomgridding.py, gridutils.py – 3D grid featurization
• blocks.py, distributions.py, models.py – VAE components
• generating.py, molfitting.py, molecules.py – generative inference
• data.py, trainutils.py, lightning.py – data loaders & PyTorch Lightning integration
• visutils.py, molfiles.py – visualization & file I/O
• Config directory: YAML templates for train, generate, optimize workflows
• Data directory: example protein/ligand sets & filters
• External files: requirements.txt, setup.py
README for additional insights
• The README confirms use of PDBBind & ZINC datasets, RDKit/OpenBabel for chemistry, PyTorch Lightning for training, and QuickVina2 for docking. It describes four main workflows: preprocessing, training, sampling, and optimization.