How thinng/graphdta works
Project type and purpose
• This is a machine‐learning research project focused on drug–target affinity prediction. It extends DeepDTA by representing small molecules as graphs and training graph neural network (GNN) models (GIN, GAT, GAT_GCN, GCN) using PyTorch Geometric.
• The README describes data preparation (create_data.py), dataset utilities (utils.py), two training regimes (training.py and training_validation.py), and four model classes under models/.
File structure
• Top‐level files:
• create_data.py: transforms raw Davis/KIBA data into PyTorch‐format datasets
• training.py & training_validation.py: orchestrate training & evaluation
• utils.py: dataset classes and performance metrics
• supplement.pdf: research supplement
• data/
• Two raw datasets (davis, kiba) with subfolders for Y (labels), proteins.txt, ligands_can.txt, and fold definitions
• After running create_data.py, data/processed/ will contain .pt train/test splits
• models/: four GNN implementations (ginconv.py, gat.py, gat_gcn.py, gcn.py)
• No dedicated frontend or service layer—this is a command‐line model training pipeline.
README for additional insights
• Clearly delineates three stages: (1) data conversion, (2) training on test‐set‐selected best MSE, (3) training with validation split and final test.
• Dependencies: PyTorch Geometric and RDKit for graph construction from SMILES.
Elements
• Clear labels for each component (e.g., “create_data.py”, “GINConvNet”)
• Orange for model classes
• Purple for external dependencies
Detailed checklist
• Data layer: raw data files, processed .pt files
• Preprocessing module: create_data.py, utils.py (dataset loader)
• Modeling: four GNN modules with a generic “Model Interface” annotation
• Training controllers: training.py vs. training_validation.py (annotate difference in selection criterion)
• Outputs: model and result files
• External: RDKit, PyG libraries