How layne-huang/pmdm works
Main Components
• Data Preparation
• Pocket & ligand extraction scripts (split_pocket_ligand.py, sample_*.py)
• Configuration files (configs/*.yml, mol.yml)
• Dataset loaders and transforms (utils/data.py, utils/datasets/, utils/transforms.py)
• Model Core
• Encoders (models/encoders/*.py: attention, cftfm, egnn, gin, schnet, etc.)
• Diffusion modules (models/diffusion.py, models/epsnet/diffusion.py)
• Geometry & common utilities (models/geometry.py, models/common.py)
• Training & Inference Orchestration
• Training scripts (train.py for single GPU, train_ddp_op.py for multi-GPU/DistributedDataParallel)
• Sampling scripts (sample.py, sample_batch.py, sample_for_pdb.py, sample_frag.py, sample_linker.py)
• Evaluation & Docking
• Evaluation metrics & processing (evaluation/*.py: docking, scoring, similarity, property_cal, MM_PBSA, etc.)
• External docking tool (QuickVina2 via evaluation/env_adt.yml and prepare_receptor4.py)
• Storage & Artifacts
• Pre-trained model checkpoints (downloaded via Zenodo)
• Generated molecule outputs and intermediate files (SDF, pdbqt)
• Logging and metrics (fpscores.pkl.gz, evaluation outputs)
Relationships and Data Flows
• Data prep → Dataset loader: raw PDB/SDF files → utils.readers → in-memory tensors
• Config loader → training/inference scripts: YAML → hyperparameters
• Training script → Model core: forward/backward through encoder + diffusion → checkpoint writes
• Sampling script → Model core: inference pass → generated 3D coordinates → write SDF/pdb
• Generated molecules → Evaluation module → compute docking (QuickVina2) & scoring → metrics output
• CLI scripts orchestrate all steps; they import from utils, models, evaluation.
Architectural Patterns & Principles
• Modular decomposition: separate directories for configs, models, utils, evaluation
• Layered design:
– CLI/entry points (train.py, sample_*.py)
– Data abstraction (utils, configs)
– Model definition (models/)
– Evaluation & external integrations (evaluation/)
• Scalability: support single-GPU and distributed training (train_ddp_op.py)
• Extensibility: plug-in style encoders, flexible sampling types via command-line flags.